Deep Learning-Derived Myocardial Strain.

Area of Science:

• Cardiology
• Medical Imaging
• Artificial Intelligence

Background:

• Echocardiographic strain measurements, particularly global longitudinal strain (GLS), are operator-dependent and exhibit significant intervendor variability.
• This variability limits the utility of retrospective research and patient analysis using standard echocardiography.
• An automated, open-source, and vendor-agnostic method is needed to improve GLS measurement accuracy and consistency.

Purpose of the Study:

• To develop and validate an automated deep learning strain (DLS) analysis pipeline.
• To assess the performance of the DLS pipeline across diverse patient populations and imaging conditions.
• To establish DLS as a reliable and reproducible tool for GLS assessment.

Main Methods:

• Assessed interobserver and intervendor variation in traditional GLS measurements.
• Simulated the impact of contour variations on speckle-tracking strain.
• Developed the DLS pipeline using semantic segmentation from EchoNet-Dynamic to calculate longitudinal strain from endocardial contours.
• Validated DLS against traditional GLS on an external dataset and applied it to populations with cardiac hypertrophy and amyloidosis.

Main Results:

• Traditional GLS showed substantial intervendor (ICC=0.29) and interobserver (ICC=0.63) variability.
• Simulated contour variations introduced significant errors in speckle-tracking measurements.
• External validation demonstrated moderate agreement between DLS and 2D GLS (ICC=0.56, bias=-3.31%).
• DLS showed significant differences in populations with cardiac hypertrophy and moderate agreement in cardiac amyloidosis (ICC=0.64, bias=0.57%).

Conclusions:

• The open-source DLS pipeline offers reduced variation compared to manual GLS measurements.
• DLS provides rapid, consistent, and vendor-agnostic quantitative results.
• The method is publicly available and robust across various imaging qualities, suitable for research and clinical applications.