Researchers at the University of Michigan have developed an AI model that can analyze brain MR scans and generate diagnostic assessments within seconds, according to a study published in Nature Biomedical Engineering.

The model, called Prima, was trained to identify more than 50 neurological conditions and to estimate how urgently patients may need treatment. In testing, it achieved diagnostic accuracy of up to 97.5% across a broad range of neurological disorders, the researchers reported.

Unlike many prior AI tools designed for narrow imaging tasks, Prima was trained on a large and diverse data set. The team used more than 200,000 MR studies comprising 5.6 million imaging sequences collected over decades at University of Michigan Health. Clinical histories and physicians’ stated reasons for ordering scans were also incorporated into the training process.
Dr. Todd Hollon, a neurosurgeon at University of Michigan Health and senior author of the study, said the system is designed to function more like a practicing radiologist by integrating imaging data with clinical context. In addition to suggesting likely diagnoses, the model can flag cases that may require immediate attention, such as suspected strokes or brain hemorrhages, and indicate which subspecialty clinician should be notified.

The research team evaluated Prima on more than 30,000 MR studies over a one-year period. According to the study, the model outperformed other state-of-the-art AI systems on overall diagnostic performance and triage prioritization.

Researchers say the technology could help address growing demand for neuroimaging services, particularly as MR volumes increase faster than the supply of trained neuroradiologists. Delays in interpreting scans can stretch from hours to days, depending on the setting and available resources.

Yiwei Lyu, M.S., a co-first author and postdoctoral fellow in computer science and engineering at the university, said the findings suggest that rapid analysis does not necessarily require sacrificing accuracy.

The authors caution that the work represents an early stage of evaluation. Future research will focus on integrating additional electronic medical record data and testing the model in broader clinical environments. The team also notes that similar approaches could eventually be applied to other imaging modalities, including mammography, chest X-rays and ultrasound.

The study was supported in part by the National Institutes of Health and several private foundations.

Back to HCB News

Artificial Intelligence Homepage