Myocardial Function Imaging in Echocardiography Using Deep Learning.

Area of Science:

• Cardiovascular Imaging
• Medical Technology
• Artificial Intelligence in Medicine

Background:

• Echocardiography deformation imaging offers superior diagnostic and prognostic value over traditional measures like ejection fraction.
• Despite proven efficacy, widespread clinical adoption of echocardiographic strain imaging is hindered by concerns regarding practical robustness and inter-vendor variability.

Purpose of the Study:

• To develop a novel deep learning framework for automated motion estimation in echocardiography.
• To fully automate myocardial function imaging, addressing limitations of current clinical practices.

Main Methods:

• A PWC-Net architecture was employed for motion estimation, achieving high accuracy on simulated data.
• A fully automated pipeline integrating cardiac view classification, event detection, myocardial segmentation, and motion estimation was developed.
• The framework incorporated image augmentations to enhance adaptability to common echocardiographic artifacts.

Main Results:

• The motion estimator achieved an average endpoint error of (0.06±0.04) mm per frame on simulated data, comparable to state-of-the-art methods.
• The automated pipeline demonstrated promising results in estimating left ventricular longitudinal strain in vivo, with a mean deviation of (-0.7±1.6)% compared to a commercial solution in 30 patients.
• The deep learning approach showed unique adaptability to image artifacts like signal dropouts.

Conclusions:

• Learning-based motion estimation provides a robust and adaptable solution for myocardial function imaging.
• The developed framework has the potential to facilitate the extended clinical use of echocardiographic strain imaging, improving patient diagnosis and prognosis.