Revealing Unforeseen Diagnostic Image Features With Deep Learning by Detecting Cardiovascular Diseases From Apical...

Area of Science:

• Cardiology
• Medical Imaging
• Artificial Intelligence

Background:

• Advancements in portable ultrasound and automated acquisition enable accessible cardiovascular disease diagnosis.
• Automated interpretation methods using limited views can enhance diagnostic accessibility.
• Deep learning offers a novel approach to cardiovascular disease detection from echocardiograms.

Purpose of the Study:

• To develop a deep learning method for automated detection of impaired left ventricular (LV) function and aortic valve (AV) regurgitation.
• To identify critical anatomical structures and temporal frames for AI-based cardiovascular disease classification.
• To assess the feasibility of 3D convolutional neural networks for echocardiogram analysis.

Main Methods:

• Extracted apical 4-chamber ultrasound cineloops from 3554 echocardiograms.
• Trained two separate convolutional neural networks for impaired LV function and AV regurgitation detection.
• Conducted feature importance analyses to understand model decision-making.

Main Results:

• Achieved 86% accuracy for impaired LV function detection.
• Achieved 83% accuracy for AV regurgitation detection.
• Identified LV myocardium and mitral valve for LV function; mitral valve anterior leaflet tip during opening for AV regurgitation.

Conclusions:

• Demonstrated feasibility of 3D CNNs for detecting impaired LV function and AV regurgitation from ultrasound.
• Deep learning can detect cardiovascular diseases using novel image features.
• AI-driven analysis may reveal unforeseen diagnostic features in echocardiograms.