Abstract:
Heart failure (HF) is a prevalent and life-threatening medical condition afflicting millions of people worldwide. A critical parameter in assessing cardiac function and detecting HF is the left ventricular ejection fraction (LVEF), which measures the amount of blood expelled from the left ventricle (LV) during each contraction, indicative of the heart's ability to efficiently circulate oxygen-rich blood throughout the body. While echocardiography has traditionally served as the primary imaging modality for evaluating LVEF due to its accessibility and cost-effectiveness, recent advancements in cardiac magnetic resonance have positioned it as an invaluable tool, particularly for detecting heart failure with preserved ejection fraction (EF). However, it's worth noting that CMR is financially more burdensome compared to echocardiography. This study leverages the EchoNet-Dynamic dataset, comprising 10,030 echocardiographic videos with annotations of LV coordinates. We have established a well-structured data preprocessing pipeline to extract frames and associated coordinates, ensuring the dataset's suitability for deep learning (DL) models. Our proposed model architecture incorporates pretrained transfer learning (TL) models optimized for localizing LV boundaries. Through the application of convolutional neural network (CNN) regression-type models to predict coordinates, we demonstrate a novel volume tracing method that may alleviate the limitations associated with segmentation-based approaches. Our findings underscore the potential of deep learning to enhance the precision and efficiency of cardiac analysis. The methodology we have developed equips healthcare practitioners with timely insights to inform clinical decision-making. © 2023 IEEE.
