Real-Time Automatic Ejection Fraction and Foreshortening Detection Using Deep Learning.

Area of Science:

• Cardiology
• Medical Imaging
• Artificial Intelligence

Background:

• 2-D echocardiography measurements of left ventricular (LV) volume and ejection fraction (EF) suffer from high uncertainty.
• Interobserver variability and ultrasound acquisition errors, like apical foreshortening, contribute to measurement inaccuracies.

Purpose of the Study:

• To develop a real-time, fully automated method for EF measurement and apical foreshortening detection in 2-D echocardiography.
• To reduce measurement errors and time in clinical echocardiography analysis.

Main Methods:

• Utilized deep learning components including view classification, cardiac cycle timing, segmentation, and landmark extraction.
• Trained neural networks on a dataset of 500 patients and evaluated on a separate dataset of 100 patients.
• Quantified apical foreshortening impact using 3-D ultrasound.

Main Results:

• The proposed method accurately detects and quantifies apical foreshortening, which significantly affects EF.
• Automatic EF measurements showed a bias and standard deviation of -3.6 ± 8.1%, with a mean absolute difference of 7.2%.
• These results are comparable to interobserver variability and related studies.

Conclusions:

• The automated real-time pipeline provides continuous acquisition and measurement without user interaction.
• The method has the potential to significantly reduce analysis time and measurement errors due to foreshortening.
• Enables quantitative LV volume measurements in routine echocardiography settings.