Two revolutions merging. Magnetic resonance imaging (MRIThis was a giant leap forward for health, allowing us, at the end of the last century, to see the brain in action for the first time. This technique, which earned Peter Mansfield and Paul Lauterbur the Nobel Prize in Medicine in 2003, opened a window onto the secrets of our most mysterious organ, propelling neuroscience into a new era.
However, some twenty years later, a a new revolution is transforming health Artificial intelligence. It was only a matter of time before these two technological breakthroughs converged. And now they have, thanks to a specialized AI developed by researchers at the University of Michigan in the United States. Their approach, which automates the detection of neurological problems during MRI scans, was presented on February 6, 2026, in the journal Nature Biomedical Engineering.
An AI that combines images and text to optimize brain analysis
The AI in question is called Prima. It's not exactly the first artificial intelligence for MRI, but it is the first fully autonomous one. Previously, other AIs had to be trained with images annotated by an expert to teach the model to recognize anomalies. Unlike Prima, which functions more like language models (LLMs) such as ChatGPT, which require a large amount of data but do the work of "understanding" that data themselves.
In addition to analyzing MRI images, Prima can also analyze text to link a patient's brain scans to their medical history and refine the diagnosis. Prima functions like a radiologist, in that it integrates the patient's medical information and image data to gain a better understanding of their health., explains in a press release the author of the study, Samir Harake. This allows it to achieve better performance across a wide range of predictive tasks.
AI correctly diagnoses patients with a success rate greater than 90% (%)
Researchers trained Prima with data from over 170,000 patients who underwent head MRIs at their university hospital before 2023 (for a total of 5.6 million MRI sequences). Each MRI was linked to a summary of the corresponding radiology report, allowing the AI to learn to interpret the images independently. Through this training, Prima was able to correctly describe the images with a success rate of 94% (%), enabling it to diagnose 52 neurological abnormalities, such as vascular disorders, inflammation, and infections in the brain.
Then, the scientists tested Prima with all new patients admitted to the hospital between June 2023 and June 2024 (for a total of 29,435 patients). The AI was able to correctly diagnose each medical case with a success rate of 90% using only brain images, and 92% using both images and the patient's medical history. For some cases, this diagnostic rate was particularly high: for example, for glial tumors (which affect glial cells, located like neurons in the brain), the success rate reached 99.7%!
Prima improves patient care
This rapid diagnostic capability could be very useful for triaging patients and optimizing their medical pathway. Researchers have shown that Prima effectively identifies the most serious cases quickly, allowing for the prioritization of their care. Furthermore, it correctly suggests which service each patient should be directed to based on their condition: for example, a person with multiple sclerosis (an autoimmune disease) should be treated by a neuro-immunology specialist.
However, the goal of this AI is not to replace the doctor, but only to support them: »In the same way that AI tools can help write an email and provide recommendations, Prima would be a co-pilot for interpreting medical imagesreassures neurosurgeon Todd Hollon, director of the study. Prima is a good example of integrating artificial intelligence models into healthcare systems. With his team, he now aims to strengthen his model by integrating additional data, such as the patient's genetics or other health information, in order to enrich its diagnostic capacity and improve care.
English 
French