Cleveland Clinic and Carnegie Mellon Publish AI That Reads Cardiac MRI With Near-Clinical Accuracy

Learning from radiology reports rather than manual labels, the research-stage model shows strong performance but needs broader validation before routine clinical use.

Overview

Researchers published the CMR-CLIP model in Nature Communications on Thursday, May 21, 2026, describing a vision-language system that links radiology reports to time-sequence MRI images to learn without manual image labels.
The model was trained on more than 1 million images from over 10,000–13,000 de-identified CMR studies collected at Cleveland Clinic over more than a decade.
In internal and limited external tests the system outperformed general-purpose models by up to about 35% and reached very high diagnostic accuracies for multiple conditions, with some tasks reporting rates as high as about 99%.
CMR-CLIP can retrieve similar cases using natural-language prompts and operate in zero-shot settings where it identifies conditions without being explicitly trained on those labels.
Authors say the work points to future uses such as automated report drafting and decision support, but they emphasize the need for wider external validation, workflow integration, and regulatory review before clinical deployment.