Heart disease is the leading cause of adult mortality worldwide, and the diagnosis and management of cardiovascular disease has become a global health priority. An echocardiogram, or heart ultrasound, is one of the most commonly used imaging tools used by doctors to diagnose a variety of heart diseases and conditions.
Most standard echocardiograms provide a two-dimensional visual image (2D) of the three-dimensional (3D) heart anatomy. These echocardiograms often capture hundreds of 2D slices, or views, of the beating heart, allowing doctors to make a clinical assessment of the heart’s function and structure.
To improve diagnostic accuracy for heart disease, researchers at the University of California, San Francisco set out to determine whether a type of AI algorithm, deep neural networks (DNNs), could be redesigned to better capture complex 3D anatomy and physiology from multiple image views simultaneously. They developed a new “multi-view” DNN structure, or architecture, that allows them to extract information from multiple image views at once, rather than current approaches that only use a single view. This architecture was then used to train a demonstration DNN to detect three cardiovascular disease conditions: left and right ventricular abnormalities, diastolic dysfunction, and valvular regurgitation.
In a study published on March 17, Nature cardiovascular researchThe researchers compared the performance of DNNs that analyzed single-view and multiple-view echocardiogram data from UCSF and the Montreal Heart Institute. They found that DNNs trained on multiple views had improved diagnostic accuracy compared to DNNs trained on a single view, demonstrating that AI models that simultaneously combined information from multiple image views better captured these cardiac pathology.
Until now, AI has primarily been used to analyze one 2D view at a time from an image or video, which has limited the ability of AI algorithms to learn disease-related information between views. DNN architectures that can integrate information across multiple high-resolution views are an important step toward maximizing AI performance in medical image processing. For echocardiography, information from a single view only tells part of the story, so most diagnoses require consideration of information from multiple views. ”
Geoffrey Tison, MD, MPH, senior study author, cardiologist, and co-director of the UCSF Biosignals Research Center
For example, when assessing left ventricular (LV) size or function, an echocardiogram view that shows all the ventricles of the heart at once (A4c) best captures certain left ventricular walls (infero-septal and anterolateral walls), whereas a separate vertical echo view (A2c) captures other important walls (anterior and inferior walls). The function of the left ventricular wall often appears perfectly normal from one perspective, but has significant dysfunction from another perspective. For the echocardiographic tasks investigated by the researchers, such as identifying left and right ventricular abnormalities and diastolic dysfunction, their results suggest that multi-view DNNs are likely to learn interrelated information between features in each view to achieve higher overall performance.
Our multi-view neural network architecture is explicitly designed to allow the model to learn complex relationships between information in multiple image views. Although we found that this approach improves the performance of diagnostic tasks in echocardiography, this new AI architecture can also be applied to other medical imaging modalities where multiple views contain complementary information. ”
Dr. Joshua Barrios, lead author of the study, assistant professor, UCSF Department of Cardiology
The researchers also found that averaging the predictions of three single-view DNNs can improve performance over single-view DNNs while reducing computational cost, making it a viable alternative to training multi-view DNNs. However, in comparison, multi-view DNN provided the strongest performance. They suggest that future research should investigate how multi-view DNN architectures can aid other medical tasks and imaging modalities.
sauce:
University of California San Francisco Medical Center
Reference magazines:
Barrios, J.P.; others. (2026). Multi-view deep learning improves detection of major heart conditions from echocardiography.Nature cardiovascular research. DOI: 10.1038/s44161-026-00786-7. https://www.nature.com/articles/s44161-026-00786-7
