Sunday, November 29 | 2:30 p.m.- 2:40 p.m. | S5-SSCA02-1 | Room E353C
AI could improve cardiac MRI exam efficiency, standardize interpretation, and expand patient access.
Is it possible to use AI for point-of-acquisition evaluation of heart structure, function, and tissue characterization on MRI? Definitely, according to researchers from the Mayo Clinic.
Presenter Omer Demirel, PhD, and colleagues developed an on-scanner cardiac MR foundation model that allows for multi-task analysis and real-time report generation. The algorithm, called ORION-CMR (On-scanner Reporting with Intelligent fOundatioN Model), is an agentic framework trained on 9,258 cardiac MR studies (consisting of 12.9 million images) and used directly on the MR scanner console. It can perform short-axis segmentation with ventricular parameter extraction; disease classification; late gadolinium enhancement (LGE) classification and segmentation; T1 mapping with American Heart Association-based segmentation; and aortic flow analysis.
Demirel's team found the following:
· Short-axis segmentation conducted by the algorithm showed Dice similarity coefficient scores of 0.92 (left ventricle, or LV), 0.83 (myocardium), and 0.87 (right ventricle, or RV).
· LGE classification accuracy was 100% for positive cases and 80% for negative cases, with segmentation Dice of 0.85 (LV).
· Aortic flow segmentation achieved 0.95 Dice, and T1 myocardium segmentation realized a Dice score of 0.82.
· Disease classification accuracy was 78% overall (hypertrophic cardiomyopathy, 90%; dilated cardiomyopathy,100%; myocardial infarction, 70%; normal, 70%; and right ventricular disease, 60%)
Check out this session to hear how the algorithm could streamline the use of cardiac MRI in clinical practice.
