Enhancing Ejection Fraction Prediction with Graph Neural Networks

The ejection fraction (EF) of the heart is a key indicator of left ventricular function however, the use of manual measurements from echocardiographic images is time-consuming and lacks interobserver variability. Deep Learning has yielded automated solutions concerning EF assessment. Although most existing models have failed to capture the complex temporal and spatial dynamics of cardiac cycles, we have addressed this limitation by presenting a novel framework with the help of Graph Neural Networks (GNNs). We have also trained the model on the EchoNet Dynamic dataset of more than 10,000 sample echocardiographic videos. The model depicts the graph representation of the echocardiographic sequences. Significant elements of our methodology include designing the Graph Neural Networks (GNNs) for spatial feature aggregation. Given several different experiments, the proposed GNN-based model has been found to outperform other models, including those with a mean absolute error of 1.7