CMR-CLIP AI model interprets cardiac MRI scans with high accuracy

Researchers from Carnegie Mellon University and Cleveland Clinic have developed an AI system designed to interpret cardiac MRI scans without manually labeled training data. The system, called CMR-CLIP, connects moving cardiac MRI images with corresponding clinical radiology reports.

CMR-CLIP was trained on more than 13,000 patient studies, according to the announcement. The researchers said the system outperformed existing general-purpose AI models by up to 35% in testing.

The study was published May 21 in Nature Communications. David Chen, PhD, of Cleveland Clinic, served as co-principal investigator, and Ding Zhao, PhD, of Carnegie Mellon University, was also a co-principal investigator on the work.

“Cardiac MRI interpretation is highly specialized and time intensive,” Chen said in the announcement. He said systems such as CMR-CLIP could support automated screening and interpretation support, particularly where expert readers are limited.

Cardiac MRI exams can include hundreds to thousands of images across views and time points. A specialized clinician can spend at least 40 minutes reviewing and annotating an exam, according to the research team’s release.

Rather than using static images alone, CMR-CLIP represents cardiac MRI exams as moving sequences. The approach allows the system to account for heart structure and motion over time, according to the researchers.

In zero-shot testing, CMR-CLIP identified cardiac conditions by matching images to descriptive prompts, such as enlarged left ventricle. The model reached accuracy as high as 99% for certain heart conditions, according to the reported findings.

The system also showed potential for natural-language search across cardiac MRI databases. Researchers said this could help match similar cases, including rare or complex presentations.

Performance was also reported on 2 separate external datasets, including 1 from France and another from Cleveland Clinic Florida. The researchers said this suggested the model could generalize beyond a single hospital system.

“This work demonstrates that domain-specific foundation models can significantly outperform general-purpose AI systems in specialized clinical applications,” Zhao said.