UCSF Researchers Develop Multiview AI Architecture to Synchronize 2D Cardiac Ultrasounds into Accurate 3D Diagnostic Models

New research from UCSF reveals a "multiview" AI architecture that analyzes multiple heart ultrasound angles at once to boost diagnostic accuracy.

By: AXL Media

Published: Mar 24, 2026, 5:14 AM EDT

Source: Information for this report was sourced from UC San Francisco

Overcoming the Limitations of 2D Cardiac Imaging

Echocardiograms are the primary tool for diagnosing cardiovascular conditions, yet they inherently present a challenge: they capture two-dimensional slices of a three-dimensional, beating organ. While a standard exam produces hundreds of these 2D views, traditional artificial intelligence models have been limited to analyzing only one view at a time. This fragmented approach often misses critical data, as a heart wall might appear healthy in one angle but show significant dysfunction in another. To solve this, researchers at UC San Francisco have re-engineered deep neural networks (DNNs) to process multiple high-resolution imaging views at once, effectively allowing the AI to "see" the heart’s complex anatomy in a more integrated, holistic manner.

Synchronized Analysis of Ventricular and Valvular Health

The UCSF team trained their new multiview architecture to detect three specific cardiovascular conditions: ventricular abnormalities, diastolic dysfunction, and valvular regurgitation. By drawing information from several perpendicular views simultaneously—such as the four-chamber and two-chamber views—the AI can identify defects across different sections of the heart muscle. For instance, while one view might capture the inferoseptal walls, a perpendicular view is required to assess the anterior and inferior walls. Senior study author Dr. Geoffrey Tison noted that this integrated approach is essential because any single view typically only tells a fraction of the physiological story.

Superior Performance in Comparative Diagnostic Trials

In a study published on March 17, 2026, the researchers compared the diagnostic performance of their multiview DNN against traditional single-view models using data from UCSF and the Montreal Heart Institute. The results demonstrated that the multiview architecture significantly improved accuracy across all examined tasks. The AI appears to learn interrelated features between the different views, allowing it to reach a higher level of clinical assessment than human-led single-view analysis. While the researchers found that averaging the results of several single-view models also improved performance, the true multiview DNN remained the strongest performer in the trial.