An artificial intelligence model capable of reading a brain MRI and providing a diagnosis in seconds, with an accuracy of up to 97.5%. This is what was developed by a team of researchers from the University of Michigan (USA), who trained the system on vast imaging archives to intercept neurological pathologies and establish treatment priorities.
The system, called “Prima”, is described in a study published on Nature Biomedical Engineering and is proposed as a tool to reduce pressure on healthcare systems. Todd Hollon, a neurosurgeon and senior author of the research, says the model has the potential to improve diagnosis and treatment by providing rapid and accurate information in the face of ever-increasing global demand for tests.
Prima is a “vision language model” (Vlm) capable of simultaneously processing video, images and text, trained on over 200 thousand historical studies and 5.6 million radiological sequences. Unlike previous algorithms based on manually selected datasets, this system integrates scans with medical records and reasons for medical prescriptions, simulating a radiologist’s complete diagnostic approach.
During tests conducted on more than 30,000 MRIs over a year, the model outperformed other AI technologies in identifying more than 50 serious neurological diagnoses, such as strokes and hemorrhages. The system was able to flag urgent cases that required priority and automatically suggest the relevant specialist, such as a neurosurgeon or neurologist.
According to Yiwei Lyu, co-author of the study, the combination of precision and speed offered by the model can simplify clinical workflows without compromising the reliability of the diagnosis. Vikas Gulani, chair of the Department of Radiology at the US university, adds that such technology could improve access to radiology services both in large medical centers and in rural hospitals with limited resources.
The authors point out that the research is still in an early evaluation phase, but they plan to expand the system’s capabilities by integrating additional data from electronic health records. Hollon describes the technology as a possible “co-pilot” for medical interpretation, hypothesizing a future application also to other diagnostic modalities such as mammograms, chest x-rays and ultrasounds.
Artificial intelligence in neuroimaging, between precision and timeliness
The integration of artificial intelligence systems in diagnostic imaging, particularly in brain MRI, represents a significant evolution in modern clinical practice. These tools were created to respond to two fundamental critical issues of global healthcare systems: the growing volume of radiological tests and the need to reduce reporting times for time-dependent pathologies. The ability to analyze large amounts of data in seconds it is not just a question of speed, but of efficiency in the clinical triage process.
The model developed at the University of Michigan stands out for its use of a multimodal approach, integrating the visual data of the scans with the patient’s clinical information. In traditional medical practice, the radiologist never interprets an image in a vacuum, but correlates it with the symptoms, medical history and diagnostic suspicion formulated by the treating doctor. The alignment of artificial intelligence with this working methodwhich combines vision and language, allows for a more accurate evaluation than models that only analyze image pixels.
Clinical decision support mechanisms
From a technical perspective, innovation lies in the nature of these defined systems vision language models. These algorithms are trained not only to recognize structural abnormalities, such as masses or vascular lesions, but also to interpret the context in which the exam was requested. This process allows you to generate a summary that can quickly direct the patient to the most suitable specialist, be it a neurologist or a neurosurgeon, optimizing treatment paths within hospitals.
A crucial aspect concerns the management of neurological emergencies, such as stroke or intracranial hemorrhages, where every minute saved can directly influence the patient’s prognosis and functional recovery. An automated warning system it can act as a safety filter, immediately alerting medical staff to scans that show signs of extreme severity, preventing them from remaining in the reporting queue alongside less urgent routine tests.
Limits, opportunities and scientific consensus
Despite the reported high diagnostic accuracy, it is essential to maintain a correct clinical perspective on the role of these technologies. The current scientific consensus does not see artificial intelligence as a replacement for the radiologist, but rather as a co-pilot capable of augmenting human capabilities. The final diagnosis remains a complex medical act that requires a synthesis of bioethical, clinical and instrumental elements which, at present, the algorithm can only support and not entirely replace.
There are still margins of uncertainty related to the so-called generalizability of the models, or their ability to maintain the same effectiveness on populations different from those on which they were trained. Future research will need to confirm whether these performances remain constant in heterogeneous clinical contexts and whether systematic integration into electronic health records leads to real improvements in long-term patient outcomes. The potential for scalability towards other imaging modalities, such as ultrasound or chest x-rays, however, suggests that we are only at the beginning of a structural transformation of the diagnostic workflow.
Source: adnkronos
date: 2026-02-08 00:41:00
Keep reading