Multiview deep learning improves detection of major cardiac conditions from echocardiography.

Multiview DNNs improved discrimination compared to single-view DNNs across all primary detection tasks.

• Area under the receiver operating characteristic curve (AUC) improved by 0.06-0.09 compared to DNNs trained on any single view.
• Three primary demonstration tasks were evaluated: detecting any left or right ventricular abnormality, diastolic dysfunction, and substantial valvular regurgitation.
• The multiview DNN architecture combines information from multiple video views simultaneously within a single model.

A deep neural network architecture was developed that integrates multiple two-dimensional echocardiographic video views simultaneously.

• The architecture is designed to combine information from multiple video views simultaneously rather than analyzing single views independently.
• The approach addresses the limitation that most AI models analyze two-dimensional data despite medical imaging capturing multiple 2D views of 3D anatomic structures.
• The multiview DNN approach was applied using a single AI model rather than ensemble or separate models per view.

The multiview DNN approach was validated using echocardiogram data from two independent institutions.

• Data were sourced from the University of California, San Francisco (UCSF) and the Montreal Heart Institute.
• Use of two geographically and institutionally distinct datasets supports generalizability of the multiview approach.
• Tasks included detection of ventricular abnormalities (left and right), diastolic dysfunction, and substantial valvular regurgitation.

Integrating multiple imaging views better captures complex cardiac anatomy and physiology for certain diagnostic tasks.

• The authors conclude that multiview models can 'better capture complex anatomy and physiology for certain tasks.'
• The findings 'underscore the value of a multiview paradigm for AI in medical imaging.'
• The improvement was demonstrated specifically for conditions where multiple anatomic windows provide complementary diagnostic information.