Prima AI: When a Brain MRI Takes Seconds, Not Days

Prima AI: When a Brain MRI Takes Seconds, Not Days

The Waiting Room Problem

Imagine this: you've just had a brain MRI. The machine has finished its loud, claustrophobic symphony of clicks and hums. You get up, change back into your clothes, and then... you wait. Hours. Sometimes days. In rural hospitals or overstretched health systems, it can take even longer before a radiologist reads your scan and your doctor calls with results.

For most people, that wait is just stressful. For someone having a stroke, it could be the difference between recovery and permanent damage.

This is the reality of medical imaging today. Demand for radiology services continues to outpace the number of specialists available to interpret scans — a gap projected to widen through 2055. The images are captured quickly, but the human expertise to read them is a bottleneck that affects millions of patients every year.

What if a model could read that MRI in seconds?

Meet Prima

In February 2026, a team at the University of Michigan published a paper in Nature Biomedical Engineering that introduced Prima — the first general-purpose vision language model designed to interpret real-world, clinical brain MRI studies. Not a narrow classifier trained on one condition. Not a prototype tested on curated research datasets. A foundation model trained on the full, messy, real-world breadth of an entire health system's neuroimaging archive.

The name is fitting. Prima, as in first. And for what it achieves, the name holds up.

Led by Dr. Todd Hollon, a neurosurgeon and AI researcher at Michigan Medicine, alongside co-first authors Samir Harake and Yiwei Lyu, the project represents a collaboration across neurosurgery, radiology, neurology, computer science, and computational medicine — with additional support from the University of Cologne in Germany.

Dr. Hollon has described Prima as something like "ChatGPT for medical imaging" — a co-pilot that integrates a patient's imaging data and medical history to produce a comprehensive diagnostic picture. But beneath that accessible tagline is a genuinely sophisticated piece of engineering.

What Prima Actually Does

Prima doesn't just look at a single image. It reads a complete brain MRI study — multiple sequences, multiple volumes — and simultaneously ingests clinical text like the patient's medical history and the physician's reason for ordering the scan.

From that, it produces:

  • Differential diagnoses across 52 radiologic conditions
  • Referral recommendations to the appropriate subspecialist
  • Triage and prioritization cues for urgent cases
  • Automatic emergency alerts for time-sensitive findings like stroke or hemorrhage

That last point deserves emphasis. Prima can flag a life-threatening scan and route it to the right specialist — a stroke neurologist, a neurosurgeon — immediately after the patient finishes imaging. Before any human has looked at it. In conditions where every minute matters, this kind of speed isn't a convenience. It's potentially life-saving.

The Numbers: 97.5% Accuracy Across 52 Diagnoses

The evaluation wasn't small. The team ran a one-year, health system-wide study at the University of Michigan covering 29,431 MRI studies across 52 radiologic diagnoses spanning the major categories of neurologic disease:

  • Neoplastic conditions — brain tumors, meningiomas
  • Vascular conditions — acute ischemic stroke, intracranial hemorrhage
  • Inflammatory conditions
  • Infectious conditions — brain abscesses
  • Developmental conditions — Chiari malformations

The results:

MetricValue
Mean diagnostic AUC (with clinical context)92.1%
Mean diagnostic AUC (MRI only)90.1%
Peak diagnostic accuracy97.5%
Explainability (Top-3 tumor region selection)98.0%

A mean AUC of 92.1% across 52 different diagnoses is remarkable. The peak of 97.5% in certain diagnostic categories is exceptional. And the fact that augmenting MRI data with clinical context (patient history, ordering reasons) improved performance demonstrates that Prima is doing something closer to what a radiologist does — it's not just reading pixels. It's reading context.

How It Works Under the Hood

For those curious about the architecture, Prima uses a three-stage training pipeline that's worth understanding.

Stage 1: Volume Tokenization

Each MRI volume is divided into small 3D patches (32×32×4 voxels). Background patches are removed. A 3D Vector Quantized-Variational Autoencoder (VQ-VAE) compresses each patch into a compact embedding vector, trained with an L1 reconstruction objective and a codebook of 8,192 entries. Random permutation of image axes during training ensures the model handles different imaging planes and orientations gracefully.

Stage 2: Hierarchical Vision Transformer

This is the core architecture — two nested transformers working in concert:

  • A Sequence ViT processes volume tokens alongside embedded text descriptions of each MRI sequence (like "AX_T2")
  • A Study ViT integrates information across all sequences to produce a single representation of the full MRI study

These are aligned using a CLIP-style contrastive objective — the MRI study embeddings are trained to match LLM-summarized radiology report embeddings. There's also a clever self-supervised patient discrimination objective that leverages the fact that different sequences from the same patient share neuroanatomic features.

Stage 3: Transfer Learning

The tokenizer and transformers are frozen. A lightweight 3-layer MLP classifier is trained on the learned features for the 52 target diagnoses. When clinical context is added through LLM-embedded text, diagnostic performance improves further — a design that mirrors how a human radiologist works: images first, clinical story second, diagnosis as the synthesis of both.

The Data Behind the Model

What makes Prima particularly compelling isn't just the architecture — it's the scale and nature of the training data.

The model was trained on UM-220K, a dataset built from every brain MRI taken since radiology digitization began at the University of Michigan:

  • 220,000+ MRI studies
  • 5.6 million 3D MRI sequences
  • 362 million 2D MRI images
  • Accompanying radiology reports and clinical histories

This is health system-scale, real-world data. Not hand-picked clean samples. Not research-only datasets. The full diversity of what walks through the doors of a major academic medical center — rare conditions, ambiguous findings, noisy images, the whole spectrum. And the team showed that performance continues to improve as data scales up, suggesting the model hasn't hit a ceiling yet.

For labeling, the team used HIPAA-compliant GPT-4 with expert-engineered prompts to annotate MRI studies across the 52 diagnoses, achieving an average annotation accuracy of 94.0% — comparable to expert human annotators. A practical demonstration that LLMs can be useful tools in the data pipeline, not just the end product.

Beyond Radiology Reports

One of the most exciting findings is that Prima's learned features transfer to tasks that don't typically appear in radiology reports:

  • Alzheimer's disease prediction
  • Autism spectrum prediction
  • Brain age estimation (mean absolute error of 5.6 years)

These results match or exceed supervised baselines on public datasets — without the extensive preprocessing those benchmarks typically require. The features Prima learns from clinical neuroimaging generalize to neuroscience research questions that the model was never explicitly trained for.

That's the promise of foundation models in medicine. Train once on broad, real-world data, and the learned representations unlock applications the original researchers didn't anticipate.

Why This Matters: The Bigger Picture

Let me put the significance in perspective with a few realities of modern radiology:

The workforce gap is real. Demand for imaging continues to grow, while the number of radiologists doesn't keep pace. This isn't a future problem — it's happening now.

Geography determines access. A patient at a major academic medical center might get their MRI read the same day. A patient at a rural hospital might wait days. Same scan, same urgency, different outcome based on where they live.

Fatigue causes errors. Radiologists are human. A radiologist who has read 80 scans in a day will miss things that the same radiologist would catch fresh in the morning. Burnout in radiology is well-documented and rising.

Speed saves lives. For stroke, hemorrhage, and other acute conditions, diagnostic delay directly correlates with worse outcomes. The "golden hour" isn't a metaphor — it's a clinical reality.

Prima doesn't replace the radiologist. It gives them a co-pilot. It flags the urgent scans. It pre-reads the routine ones. It ensures that a life-threatening finding at 3 AM on a Sunday doesn't sit in a queue behind 40 other studies.

As Dr. Vikas Gulani, Chair of Radiology at Michigan Medicine, noted: "Whether you are receiving a scan at a larger health system that is facing increasing volume or a rural hospital with limited resources, innovative technologies are needed to improve access to radiology services."

Fairness and Equity

A detail that's easy to overlook but matters enormously: the team evaluated Prima for algorithmic fairness across demographic groups. In a field where AI systems have historically perpetuated or amplified health disparities, this is a thoughtful and necessary step.

The research also highlights that AI-assisted triage could help mitigate existing biases in health systems — like the documented pattern of prolonged turnaround times for underserved populations. If the AI flags urgency based on what's in the scan rather than who the patient is, it could actually improve equity rather than erode it.

What Prima Is Not (Yet)

It's important to be clear about what Prima is today:

  • It is not FDA-cleared or commercially available. It's an investigational research tool.
  • It has been evaluated at one health system (though a very large one). Multi-site, multi-population validation is the next step.
  • It focuses on brain MRI. Expansion to other imaging modalities — mammograms, chest X-rays, ultrasounds — is future work.
  • The model parameters are publicly available under an MIT license on GitHub, which is a commendable commitment to open science.

The researchers are transparent about these limitations. Future studies will test Prima across more hospitals, more diverse patient populations, and under real-world clinical deployment conditions. This is the responsible way to develop medical AI — rigorously, transparently, and with an honest acknowledgment of what remains to be proven.

A Foundation for What Comes Next

Dr. Hollon's lab at Michigan — the Machine Learning in Neurosurgery (MLiNS) Lab — has built a track record that gives confidence Prima isn't a one-off result. Their previous work includes FastGlioma, a visual foundation model for detecting brain tumor infiltration during surgery (published in Nature), and rapid AI-based molecular classification systems tested in prospective, multicenter, international clinical trials.

Prima fits into a larger trajectory: foundation models trained on real clinical data, validated at health system scale, and released openly for the research community. Each project builds on the last. Each one gets closer to clinical integration.

Looking Forward

We're at an interesting moment in medical AI. The field has moved past the "AI will replace doctors" hype cycle and into something more grounded: AI as a tool that makes doctors faster, more consistent, and more available — especially where they're needed most.

Prima is a beautiful example of this. It doesn't claim to replace the radiologist. It claims to read the scan in seconds so the radiologist can focus where their expertise matters most. It claims to flag the emergency at 3 AM so the right specialist gets paged before it's too late. It claims to bring diagnostic speed to the rural hospital that doesn't have a neuroradiologist on staff.

And with 97.5% peak accuracy across a year of real-world evaluation on nearly 30,000 scans — those aren't just claims. They're evidence.

The wait between the MRI machine and the diagnosis has always been a gap where outcomes are lost, anxiety builds, and health equity frays. Prima is a step toward closing that gap. Not perfectly. Not completely. But meaningfully.

And sometimes, a few seconds can change everything.

References and Resources:

If you or someone you know is waiting on imaging results and facing delays, know that researchers around the world are working to make that wait shorter, fairer, and safer. The future of medical imaging isn't about replacing the people who care for us — it's about giving them better tools to do it.