University Of California San Francisco Engineers Multiview Deep Neural Networks To Maximize Accuracy In Heart Disease Diagnosis

UCSF researchers develop a new AI architecture that analyzes multiple echocardiogram views simultaneously to improve the diagnosis of heart conditions.

By: AXL Media

Published: Mar 17, 2026, 12:17 PM EDT

Source: Information for this report was sourced from University of California San Francisco Medical Center

Revolutionizing Cardiac Diagnostics Through Multidimensional AI

Heart disease remains the leading cause of mortality among adults globally, placing an immense burden on healthcare systems to provide rapid and accurate diagnoses. While the echocardiogram is a staple imaging tool for physicians, most standard procedures generate two-dimensional slices of the heart's three-dimensional structure. This requires clinicians to mentally piece together hundreds of views to assess overall function. Researchers at UC San Francisco have sought to bridge this gap by redesigning deep neural networks (DNNs) to capture complex anatomy and physiology from multiple imaging views simultaneously, rather than processing them in isolation.

The Structural Evolution of Deep Neural Network Architecture

The core innovation behind the UCSF study is a new "multiview" DNN architecture specifically engineered to draw information from several perspectives at once. Traditionally, AI in medical imaging has been limited to analyzing one 2D video or image at a time, which senior study author Geoffrey Tison notes restricts the algorithm’s ability to learn "disease-relevant information" that exists between different views. By training these demonstration DNNs on data from both UCSF and the Montreal Heart Institute, the team successfully improved the detection of left and right ventricular abnormalities, diastolic dysfunction, and valvular regurgitation.

Integrating Complementary Perspectives for Precision Medicine

The necessity for a multiview approach is rooted in the inherent limitations of standard cardiac imaging. For instance, evaluating the size or function of the left ventricle requires viewing different walls of the heart, such as the inferoseptal and anterolateral walls. While one specific view may capture these sections perfectly, a perpendicular view is required to assess the anterior and inferior walls. According to study first author Joshua Barrios, the multiview neural network is explicitly designed to learn the complex, interrelated features between these complementary angles, ensuring that dysfunctions that appear normal in one view are identified in another.