Real-Time Standard View Classification in Transthoracic Echocardiography Using Convolutional Neural Networks

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Cardiology

Background:

• Transthoracic echocardiography (TTE) relies on standard cardiac views for accurate heart function assessment.
• Manual classification of cardiac views can be time-consuming and prone to variability.
• Automating view classification can streamline echocardiography workflows.

Purpose of the Study:

• To develop and evaluate convolutional neural networks (CNNs) for automatic classification of cardiac views in TTE.
• To assess the performance of CNNs in real-time and their potential for advanced applications like multiplanar reformatting.

Main Methods:

• Utilized a dataset of over 7000 echocardiography videos from more than 500 patients.
• Trained CNN models to predict up to seven different cardiac views from 2-D ultrasound data.
• Explored using 3-D data for training models applied to 2-D classification.

Main Results:

• Achieved state-of-the-art classification accuracies of 98.3% on single frames and 98.9% on sequences.
• Demonstrated real-time performance with a processing time of 4.4 ± 0.3 ms per frame.
• Obtained a median deviation of 4° ± 3° from optimal orientations when using 3-D trained models for 2-D classification.

Conclusions:

• CNNs show significant potential for automating cardiac view classification in echocardiography.
• The developed models offer high accuracy and real-time performance, enhancing clinical workflow efficiency.
• CNNs can aid in automatic multiplanar reformatting and orientation guidance, improving diagnostic accuracy.