Deep learning-based automated left ventricular ejection fraction assessment using 2-D echocardiography.

Area of Science:

• Cardiology
• Medical Imaging
• Artificial Intelligence

Background:

• Deep learning (DL) shows promise for automated left ventricle (LV) ejection fraction (EF) measurement.
• Previous DL models lacked evaluation across diverse heart disease phenotypes and echocardiography systems.
• Accurate LVEF assessment is crucial for diagnosing and managing heart failure.

Purpose of the Study:

• To evaluate a novel DL algorithm (DPS-Net) for automated LVEF measurement using 2D echocardiography (2DE).
• To assess the algorithm's performance across different heart disease phenotypes and ultrasound machines.
• To compare DPS-Net's LV segmentation accuracy with existing algorithms.

Main Methods:

• Developed a DL algorithm (DPS-Net) based on U-Net architecture using 36,890 2DE frames from 340 patients.
• Applied the biplane Simpson's method for LVEF calculation.
• Tested DPS-Net on the CAMUS dataset and compared its LV segmentation performance against EchoNet-dynamic.

Main Results:

• DPS-Net demonstrated high performance in LV segmentation (Dice coefficients 0.932 and 0.928) and LVEF measurement across diverse phenotypes and systems.
• DPS-Net v2 showed superior LV segmentation compared to EchoNet-dynamic (P = 0.008).
• Achieved high diagnostic performance for heart failure detection (AUCs ranging from 0.948 to 0.974) across phenotypes like atrial fibrillation, hypertrophic cardiomyopathy, and dilated cardiomyopathy.

Conclusions:

• DPS-Net offers accurate and robust automated LVEF measurement from 2DE images, adaptable to various echocardiographic systems.
• The algorithm exhibits high diagnostic performance for identifying heart failure across different disease phenotypes.
• DPS-Net's strong performance in LV segmentation suggests broad applicability in 2DE image interpretation.