AI MRI Model Achieves 97.5% Diagnostic Accuracy

A team of University of Michigan scientists has unveiled an AI system that can automatically diagnose patients from MRI scans. This development has the potential to revolutionize the industry, providing faster and more accessibility to patients across the globe. Here’s what you need to know.

How MRI Technology Works in Brain Imaging

Magnetic Resonance Imaging relies on magnetic fields and radio waves to create in-depth 2D/3D images of organs and other vital body parts. MRI technology was born out of earlier work on magnetic resonance (NMR) done during WWII.

In 1952, Felix Bloch and Edward Purcell took the concept further, winning the Nobel Prize in 1952. However, it wasn’t until Paul Lauterbur added spatial gradients that the technology was able to create 2D images of organs and tissues. Notably, the radio frequency pulses used in MRI scans don’t produce ionizing radiation, making them ideal compared to CT scans.

Today, MRI scans are common. According to reports, there are 150-200 million scans completed yearly. Additionally, studies show that MRI use is on the rise, with most regions experiencing a 3-6%  increase in these procedures. As such, it’s no surprise to learn that MRIs are now a crucial component of the medical system.

Why Interpreting Brain MRI Scans Is Challenging

As MRI dependence increases, it sheds light on some of the drawbacks of this technology. For one, interpreting these images takes time and precision. This demand has increased amongst growing staffing shortages in the medical sector.

The results are that as MRI usage increases, it places a strain on the already flexed medical system. Studies show that the strain differs whether it’s a large hospital dealing with an overflow of patients or a smaller facility with no access to professionals capable of interpreting MRI images. However, the results are the same – patients suffer increased risks.

Clinical Risks of Delayed MRI Interpretation

The problem is that many of the neurological diseases that MRIs help to detect require immediate attention. This requirement means that there is a balance between timely diagnosis and improved outcomes that often creates missed deadlines.

For example, brain hemorrhages and strokes are prime examples of medical conditions that MRI scans can detect and require immediate medical attention. However, these conditions are notoriously difficult to spot, requiring professional viewing to accomplish the task.

Regions of the Brain

Neuroimaging is a valuable tool for evaluating patients with neurological diseases. Consequently, global demand for MRI scans has risen steadily, placing substantial strain on health systems, prolonging turnaround times, and intensifying physician burnout.

University of Michigan Develops AI MRI Model

The “Learning neuroimaging models from health system-scale data¹” study published in Nature Biomedical Engineering introduces a purpose-built AI algorithm that can interpret MRI images in seconds. As such, it has the potential to revolutionize healthcare moving forward.

Prima: A Vision-Language AI Model for MRI Diagnosis

The Prima AI MRI imaging algorithm has been described by its creators as the “ChatGPT for medical imaging.” The AI system streamlines and automates many of the most complex aspects of medical image interpretation.

Prima differs from other AI MRI scanners in that it’s a vision language model (VLM), meaning that it processes text, images, and video together. This use of a hierarchical vision architecture enables the system to provide unmatched accuracy and performance.

How Prima Functions as a Diagnostic Co-Pilot

Speaking on the algorithm, the developers stated that they wanted to create a reliable co-pilot for interpreting medical imaging studies. To that extent, they succeeded, as Prima’s unique approach enables it to operate across a broad spectrum of diagnoses.

This approach is in stark contrast to its predecessors, who mainly zeroed in on a specific subsector of MRI imaging. Since Prima was trained on a much broader dataset, it can review the MRI data like a real physician. As such, Prima promises to improve everything from diagnosis to workflows.

Training the Model on 200,000+ MRI Studies

To train Prima, the scientist knew they would need to access a large trove of data. Luckily, the University of Michigan Health provided direct access to +200,000 MRI studies, which included more than 5.6 million image sequences.

The engineers then integrated the results of each study, including patient histories and the reasoning behind the scan. This data was then cross-referenced, and 50 different radiologic diagnoses were integrated, covering several neurological conditions.

How the AI Generates Differential Diagnoses

Prima was built to provide timely diagnosis. The system cross-references its massive database of MRI scans and corresponding data to identify a wide range of neurological conditions. Additionally, it takes into account each patient’s special medical history, enabling it to create a comprehensive file on each patient, alongside a diagnosis.

AI-Driven Emergency Triage System

Impressively, Prima can determine if the diagnosis requires immediate medical care. When the need for urgent care is detected, the system automatically notifies health care professionals, ensuring that there is no delay between diagnosis and care. This feature helps to save lives in cases where immediate action is required.

Automated Specialist Notification Workflow

One of the coolest aspects of Prima is that it automatically notifies the correct subspecialist after its diagnosis. For example, if the scan returns the risk of stroke, the system knows to contact the neurosurgeon and request emergency care.

Clinical Testing and Validation of the AI Model

Testing of the Prima system began a year ago. The scientist spent this time conducting thousands of diagnoses on patients and then comparing them to the physician’s diagnosis. In total, 29,431 MRI studies were completed.

AI MRI Model Achieves 97.5% Diagnostic Performance

Swipe to scroll →

The accuracy of Prima proved the engineers were correct. The model scored 97.5%, which is far past any AI competitor’s results. Keenly, the team noted that the system was able to complete its diagnosis in seconds versus what would have taken physicians days.

The tests also revealed that the system improved the care response time significantly. The data from the scans and diagnosis was immediately available to care specialists, and those who needed the data were automatically sent the information that required priority attention, saving patients valuable time.

Key Benefits of AI-Powered MRI Interpretation

There’s a long list of benefits that this technology brings to the market. For one, it reduces the time needed to conduct and interpret MRI scans. This improved performance helps patients receive care faster, even as the medical sector deals with major staffing shortages.

Proves AI Assistance is Ideal

Another benefit of the technology is that it helps to prove how AI systems can enhance many of the current medical procedures. This system reduces the burden on healthcare professionals while improving the results. As such, it serves as a valuable example of the power of AI systems across the medical sector.

Diagnosis

Unlike some AI systems that provide no reasoning as to how they arrived at a certain conclusion, the Prima AI model offers explainable differential diagnoses. This approach enables its work to be double-checked by physicians to ensure accuracy and gain second opinions.

Real-World Applications and Deployment Outlook

There are many applications of this technology, with the obvious one being to improve the current medical practices. Notably, this technology can be applied to other imaging sectors, leading many analysts to predict that it will be applied to mammograms, chest X-rays, and ultrasounds in the coming years.

Regulatory Pathway and Commercial Timeline

This technology could hit the market in the next 5-years. There’s strong demand for anything that can improve healthcare results while reducing the workload and stress on today’s healthcare professionals.

Research Team and Institutional Backing

The AI MRI interpretation study was put forth by University of Michigan researchers. The paper was led by neurosurgeon and assistant professor of neurosurgery at the U-M Medical School, Todd Hollon, and a Machine Learning in Neurosurgery Lab data scientist, Samir Harake.

The study lists Vikas Gulani, M.D. Ph.D.,: Asadur Chowdury, M.S., Soumyanil Banerjee, M.S., Rachel Gologorsky, Shixuan Liu, Anna-Katharina Meissner, M.D., Akshay Rao, Chenhui Zhao, Akhil Kondepudi, Cheng Jiang, Xinhai Hou, Rushikesh S. Joshi, M.D., Volker Neuschmelting, M.D., Ashok Srinivasan, M.D., Dawn Kleindorfer, M.D., Brian Athey, Ph.D., Aditya Pandey, M.D., and Honglak Lee, Ph.D. as contributors. Alongside financial support from the National Institute of Neurological Disorders and Stroke of the National Institutes of Health.

Future Development of AI in Neuroimaging

The engineers acknowledge that they are still in the early stages of their work, but the results look promising. Now, their goal is to personalize the system, enabling the AI to access medical records and make diagnoses based on all available data in real time.

Investing in AI / Healthcare Integrations

The impact of AI integration into medical imaging is accelerating. Unlike diversified device manufacturers, a new generation of companies is building AI-native radiology platforms designed specifically to improve diagnostic speed, workflow efficiency, and clinical decision-making. Here’s one public company directly positioned in the AI-powered imaging space.

RadNet (RDNT)

RadNet is one of the largest outpatient diagnostic imaging providers in the United States, operating hundreds of imaging centers nationwide. Founded in 1981, the company has grown into a vertically integrated imaging platform combining radiology services with artificial intelligence through its DeepHealth subsidiary.

Rather than treating AI as an add-on, RadNet has integrated machine learning directly into imaging workflows. Through acquisitions and internal development, the company has built AI tools designed to assist radiologists in detecting breast cancer, lung nodules, neurological abnormalities, and other conditions earlier and more consistently.

A major inflection point came with the expansion of DeepHealth, RadNet’s AI division, which focuses on deploying FDA-cleared algorithms across its imaging network. By combining proprietary imaging data with AI development, RadNet benefits from both service revenue and software-enhanced diagnostic capabilities.

This vertically integrated approach provides a competitive advantage: real-world clinical data feeds algorithm refinement, while AI tools enhance throughput and diagnostic precision within its own imaging centers.

AI-Driven Imaging Platform Expansion

RadNet continues to expand its AI portfolio through partnerships and acquisitions, targeting scalable deployment of diagnostic support tools across radiology subspecialties. As reimbursement models evolve and imaging volumes rise, AI-assisted triage and detection systems may materially improve margins while reducing radiologist workload.

For investors seeking exposure to AI-powered medical imaging with direct operational leverage, RadNet represents a more focused play on radiology automation than diversified medtech conglomerates.

Latest RadNet (RDNT) News and Performance

AI Interpretation Brain Imaging | Conclusion

The AI Interpretation brain imaging study opens the door for a brighter future for patients and medical professionals. These systems will help ensure that patients receive top-notch care that is consistent and available around the clock.

It also reduces the workload on today’s already trained healthcare professionals. As such, this development could help to create a more effective healthcare system in the future.

Learn about other cool AI developments here.

References

^{1. Lyu, Y., Harake, S., Chowdury, A. et al. Learning neuroimaging models from health system-scale data. Nat. Biomed. Eng (2026). https://doi.org/10.1038/s41551-025-01608-0}