As demand for brain imaging rises and radiologists face growing backlogs, scientists are testing whether large-scale AI models can help close the gap. Researchers at the University of Michigan have developed a foundation model for neuroimaging called Prima that can interpret brain MRI studies in seconds and flag cases that require urgent care. Prima was trained on more than 220,000 MRI studies drawn from digitized radiology records from University of Michigan Health and evaluated in a year-long prospective study involving 29,431 MRI exams. The results were published earlier this month in Nature Biomedical Engineering.
Magnetic resonance imaging is widely used to evaluate neurological disease, but demand has increased faster than radiology capacity in many regions. The authors note that growing imaging volumes have increased turnaround times and added strain to clinical workflows, especially in rural and limited-resource settings. Prima was designed to operate on real-world clinical MRI studies and associated radiology reports, with the goal of supporting routine diagnostic work.
(Wangbar/Shutterstock)
“As the global demand for MRI rises and places significant strain on our physicians and health systems, our AI model has the potential to reduce burden by improving diagnosis and treatment with fast, accurate information,” said senior author Todd Hollon, M.D., a neurosurgeon at University of Michigan Health and assistant professor of neurosurgery at U-M Medical School, in an article from Noah Fromson on MichiganMedicine.org.
According to the paper, Prima achieved a mean area under the curve of 92% across 52 radiologic diagnoses, a measure of overall diagnostic accuracy across major neurologic disorders. The model outperformed other general-purpose and medical AI systems tested on the same tasks, the authors claim. In addition to diagnostic classification, Prima generates explainable differential diagnoses, assigns worklist priority for radiologists, and provides referral recommendations.
Prima is a vision language foundation model, a multimodal AI system designed to learn from both medical images and associated text. The model uses a hierarchical vision transformer architecture pretrained with a contrastive objective that aligns three-dimensional MRI volumes with paired radiology reports. MRI studies are divided into subvolumes and encoded using a vector quantized variational autoencoder, generating tokens that are passed through sequence and study level transformers to produce a representation of the full exam. In a transfer learning step, the pretrained components are frozen and a small feedforward network is trained on the learned study features to predict specific diagnoses and clinical outcomes.
(Roman Zaiets/Shutterstock)
To build the training dataset, the team queried a Michigan Health server for all cranial MRIs with associated reports. After filtering for metadata completeness, the resulting UM-220K dataset included 221,147 studies and 5.6 million imaging sequences. LLMs were used to summarize radiology reports and extract structured findings for contrastive pretraining. The authors report no statistically significant difference between model performance aligned with LLM-generated summaries and performance aligned with human expert summaries.
Prima was also evaluated for fairness across demographic groups. Intersectional analyses of sex, race, and geographic region in the prospective cohort showed minimal variation in diagnostic performance, and equalized odds analysis indicated limited disparity in true positive and false positive rates across groups. Future versions of Prima could include the ability to integrate more detailed patient information and electronic medical record data for more accurate diagnosis, according to the Michigan Medicine article. Prima could also one day be applied with other types of imaging, like mammograms and ultrasounds.
“Like the way AI tools can help draft an email or provide recommendations, Prima aims to be a co-pilot for interpreting medical imaging studies,” Hollon said. “We believe that Prima exemplifies the transformative potential of integrating health systems and AI-driven models to improve health care through innovation.”
Find the paper at this link. All code and scripts to reproduce the training and inference of Prima are available via GitHub at MLNeurosurg/Prima under an MIT license.
The post AI Model Reads Brain MRIs in Seconds, Flags Urgent Cases appeared first on AIwire.
