Researchers at the University of Michigan have developed an advanced artificial intelligence (AI) model named Prima that can interpret brain magnetic resonance imaging (MRI) scans in a matter of seconds, according to recent studies published in Nature Biomedical Engineering. The model demonstrated a diagnostic accuracy of up to 97.5% across a broad range of neurological conditions.
The Prima AI system uses a hierarchical vision architecture that learns imaging features from both MRI scans and related clinical data. This means it doesn’t just “see” the images, it integrates a patient’s medical history and clinical context to aid diagnosis, similar to how a trained radiologist interprets scans.
Researchers trained Prima on more than 220,000 MRI studies and tens of thousands of clinical imaging sessions, enabling it to differentiate and identify over 50 neurological conditions, including strokes, brain tumors, and hemorrhages.
In testing across large real-world datasets, Prima achieved accuracy up to 97.5% in correctly diagnosing conditions seen on brain MRI scans, a level consistent with rigorous clinical evaluation metrics reported in scientific publications.
Importantly, the system can also flag cases that require urgent medical intervention, helping clinicians prioritize patients who need faster treatment. This triage capability distinguishes Prima from many earlier AI models focused solely on detection.
The developers of Prima highlight that global demand for MRI imaging is rising, creating strain on radiologists and diagnostic services. In many settings, especially in rural or underserved healthcare systems, access to expert radiology interpretation may be limited. AI models like Prima could help reduce diagnostic delays and support faster clinical decision-making.
However, researchers emphasize that AI systems are tools intended to augment human clinicians, not replace them. Integration into clinical workflows typically requires validation in real-world medical settings and continued oversight by trained physicians.
AI approaches to medical imaging are advancing rapidly. Other research initiatives are exploring deep learning methods to perform diverse tasks from brain age estimation to predicting disease outcomes using MRI data. A model called BrainIAC, for example, leverages self-supervised learning to extract multiple health indicators from MRI scans and may assist future diagnostics beyond traditional labeling systems.
While performance metrics like accuracy provide useful benchmarks, AI diagnostic systems often include additional evaluation measures such as sensitivity and specificity to fully characterize clinical reliability. Broad adoption in healthcare typically involves regulatory review, real-world clinical trials, and evaluation of how the tool performs across different populations and imaging equipment.
Reference
Lyu, Yiwei, Samir Harake, Asadur Chowdury, Soumyanil Banerjee, Rachel Gologorsky, Shixuan Liu, Anna-Katharina Meissner, Akshay Rao, Chenhui Zhao, Akhil Kondepudi, Cheng Jiang, Xinhai Hou, Rushikesh S. Joshi, Volker Neuschmelting, Ashok Srinivasan, Dawn Kleindorfer, Brian Athey, Vikas Gulani, Aditya Pandey, Honglak Lee, and Todd Hollon. “Learning Neuroimaging Models from Health System-Scale Data.” Nature Biomedical Engineering, February 6, 2026. https://doi.org/10.1038/s41551-025-01608-0.
(Rh)
