MARS-Net: Multi-Scale Attention Residual Spatiotemporal Network for Robust Left Ventricular Ejection Fraction Prediction in Echocardiography Videos

Abstract

Left ventricular ejection fraction (LVEF) is a key measure of heart pumping performance, playing a pivotal role in the ongoing management and efficacy assessment of cardiovascular disease treatments. By quantifying the percentage of blood that is pumped out of the left ventricle with each heartbeat, LVEF provides invaluable insights into the overall efficiency of the heart's function, enabling clinical professionals to make informed decisions regarding point of care and therapeutic strategies. However, accurate LVEF measurement faces challenges such as large observer variability, poor image quality, and the complexity of cardiac motion. To address these issues, a residual spatiotemporal network with a multi-scale attention mechanism is proposed for robust LVEF prediction in transthoracic echocardiographic videos, named the Multi-scale Attention Residual Spatiotemporal Network (MARS-Net). The MARS-Net excels at extracting spatiotemporal features from echocardiographic videos, accurately capturing heart dynamics and morphology while demonstrating robust performance across multi-center data. The sub-video division block is first designed to partition echocardiographic videos into smaller sub-videos, capturing key cardiac motion. The input embedding block compresses these sub-videos for efficient processing. Then the multi-scale attention residual block enhances spatiotemporal feature extraction by combining multi-scale convolutions with attention mechanisms to improve focus on important details. Finally, the output convolutional block transforms the extracted features into the final LVEF prediction, ensuring accurate measurement. Through extensive evaluations, our MARS-Net outperforms comparative deep learning models in LVEF prediction, offering exceptional promise for diagnosing heart dysfunction. Notably, it has achieved commendable results in three medical centers, underscoring its generalizability and reliability across varied clinical environments.