Researchers from the University of California, San Francisco announced on March 23 that they have developed a new artificial intelligence (AI) architecture designed to enhance the accuracy of heart disease diagnosis by analyzing multiple views from echocardiograms simultaneously.
This development is significant because heart disease remains the leading cause of adult death worldwide, and accurate diagnosis is crucial for effective treatment. Echocardiograms are commonly used imaging tools that produce two-dimensional images of the three-dimensional structure of the heart, allowing physicians to assess cardiac function and detect various conditions.
The research team created a "multiview" deep neural network (DNN) architecture that integrates information from several imaging perspectives at once. They trained demonstration DNNs with this structure to identify three cardiovascular conditions: left and right ventricular abnormalities, diastolic dysfunction, and valvular regurgitation. In a study published March 17 in Nature Cardiovascular Research, these multiview DNNs were compared against traditional single-view models using data from UCSF and the Montreal Heart Institute. The results showed improved diagnostic accuracy when multiple views were analyzed together.
"Until now, AI has primarily been used to analyze one 2D view at a time—from either images or videos—which limits an AI algorithm's ability to learn disease-relevant information between views," said Geoffrey Tison, MD, MPH, senior study author and co-director of the UCSF Center for Biosignal Research. "DNN architectures that can integrate information across multiple high-resolution views represent a significant step toward maximizing AI performance in medical imaging. In the case of echocardiography, most diagnoses necessitate considering information from more than one view because the information from any single view tells only part of the story."
The researchers explained that certain aspects of cardiac function may appear normal in one echocardiogram view but abnormal in another; therefore, combining insights across different angles provides a fuller picture for clinicians. Additionally, they found that averaging predictions from three single-view DNNs could also boost performance while being less computationally demanding than full multiview models—though true multiview DNNs still offered superior results overall.
Looking ahead, the team suggested further studies should explore how similar multiview DNN approaches might be applied to other medical tasks or types of imaging.
