AI model reads brain MRI scans in just seconds, predicting treatment urgency

A University of Michigan team has reported that a novel artificial intelligence model interpreted brain magnetic resonance imaging scans with up to 97.5 per cent accuracy across more than 50 neurological diagnoses, identifying the most urgent cases

An artificial intelligence (AI) model developed at the University of Michigan (U-M), Ann Arbor, Michigan, USA, has interpreted a brain magnetic resonance imaging (MRI) scan in seconds and has diagnosed neurological disease with accuracy of up to 97.5 per cent, according to a study that has evaluated performance across a large real-world clinical dataset. The research team reported that the same system also predicted clinical urgency allowing for rapid triage in time-critical conditions such as haemorrhage and stroke.

The senior author, Dr. Todd Hollon, a neurosurgeon at U-M Health and assistant professor of neurosurgery at the U-M Medical School, said the model could help health systems manage rising imaging demand.

“As the global demand for MRI has risen and has placed significant strain on our physicians and health systems. Our AI model has had potential to reduce burden by improving diagnosis and treatment with fast, accurate information,” Hollon said.

The system has been named Prima and his team tested it on more than 30,000 MRI studies across a single year and compared performance with other state-of-the-art AI models. The investigators reported superior diagnostic performance across more than 50 radiological diagnostic categories that covered major neurological disorders. The team also found that Prima identified those cases which required higher priority review – a capability with direct relevance for acute pathways where time is imperative.

The co-first author Yiwei Lyu, a doctoral candidate in computer science and engineering at the U-M, stressed the balance between speed and precision in neuroradiology practice.

“Accuracy has been paramount when clinicians read a brain MRI scan. But quick turnaround times have remained critical for timely diagnosis and improved outcomes,” said Lyu.

“At key steps in the process, our results have shown how Prima could improve workflows and streamline clinical care without loss of accuracy,” he added.

Prima is a vision language model, which is an AI architecture that can process images, video and text in real time. Although other groups have applied AI to neuroimaging, the U-M team argued that Prima differed in scope and training strategy. Many earlier systems relied on manually curated subsets of imaging data to complete narrow tasks, for example lesion detection or dementia risk estimation. By contrast, Hollon’s group trained Prima on the full historical MRI archive available since digitisation began at U-M Health. That training corpus included more than 200,000 studies and 5.6 million imaging sequences.

The team did not limit model input to image data with the researchers also incorporated clinical history and the referring clinician’s indication for imaging which – importantly – aimed to replicate the contextual reasoning a radiologist uses during interpretation. The co-first author Samir Harake, a data scientist and medical student in the machine learning in neurosurgery laboratory, described this design choice as central to model behaviour.

“Prima has worked like a radiologist by integrating information regarding the patient’s medical history and imaging data to produce a comprehensive understanding of their health,” Harake said.

He added that this approach enabled stronger performance across a broad range of prediction tasks rather than a single isolated output.

The reported triage function could prove especially important in settings where neuroradiology capacity is constrained. For suspected acute stroke or intracranial haemorrhage, Prima flagged urgency and could alert the most relevant specialist pathway, such as stroke neurology or neurosurgery, immediately after image acquisition. In practical terms, that means the system could support prioritisation before a full conventional reporting queue completes, with the potential to shorten time taken to make treatment decisions.

The authors positioned the work against a wider service pressure with millions of MRI examinations taking place worldwide each year and a substantial share focused on neurological disease. Demand has outpaced specialist reporting capacity in many regions which has contributed to backlog, delayed reporting and the risk of diagnostic error. In some care pathways, patients can wait days for formal results and delays can become longer in resource-limited settings.

Dr. Vikas Gulani, co-author and chair of radiology at U-M Health, said deployment context mattered as much as algorithm performance.

“Whether a person has received a scan at a larger health system that has faced increasing volume or at a rural hospital with limited resources, innovative technologies [are] necessary to improve access to radiology services,” said Gulani.

He added that multidisciplinary collaboration at U-M had produced a potentially scalable response to that structural problem.

The investigators also set clear limits on interpretation of current findings describing the present study as an early-stage evaluation rather than final proof of broad clinical effectiveness across all institutions. Future research has aimed to integrate more detailed patient-level information from electronic medical records which may improve diagnostic precision and align model inference more closely with day-to-day physician reasoning. External validation across diverse populations, scanner platforms and workflow environments will remain essential before widespread implementation.

The broader policy context remains unsettled given that health systems, regulators and professional bodies continue to grapple with an appropriate pathway of how to integrate AI into clinical practice with appropriate governance, accountability and safety monitoring and considerations.

Most tools in active clinical use still address only narrow tasks but Prima represents a shift to a more generalist model class with the potential to extend use beyond neuroimaging to other modalities, including mammography, chest radiography and ultrasound upon validation studies confirming transferability.

Hollon characterised Prima as the ‘ChatGPT of medical imaging’.

“Like the way AI tools can help draft an email or provide recommendations, Prima has aimed to act as a co-pilot for interpretation of medical imaging studies,” Hollon added.

He said the team believed the project illustrated how integrated health-system data and AI could improve diagnostic care at scale.

Taken together, the findings have suggested a credible near-term route to reduce reporting bottlenecks in neuroimaging while preserving diagnostic reliability. The critical test now will be translation from controlled evaluation through regulatory approval, to deployment into routine care, where real-world variability, clinical governance and medico-legal responsibility converge into a reliable patient benefit. 

For further reading please visit: 10.1038/s41551-025-01608-0