This paper introduces EchoCLIP, a vision-language foundation model for echocardiography, trained on a large dataset of cardiac ultrasound videos and corresponding expert interpretations. EchoCLIP demonstrates strong performance on various benchmarks for cardiac image interpretation, including assessing cardiac function, identifying implanted devices, and identifying unique patients across multiple videos. A long-context variant, EchoCLIP-R, further enhances these capabilities, enabling accurate clinical transition identification and robust image-to-text search. This research represents a significant advancement in applying foundation models to cardiovascular imaging for preliminary echocardiographic interpretation.

Matthew Christensen, Milos Vukadinovic, Neal Yuan, David Ouyang