Accurate Left Ventricle Segmentation and Ejection Fraction Estimation with Deep Learning-Based Echocardiography

Abstract

Heart failure with reduced ejection fraction is a rapidly growing public health issue with an estimated prevalence of over 37 million individuals globally. We proposed an efficient deep learning framework of automated echocardiography for predicting heart failure with reduced ejection fraction, which is most important for earlier diagnosis of heart function. The proposed deep learning framework sensibly overcame the echocardiography challenges which are domain expertise much needed for segmenting the left ventricular end-diastolic volume and end-systolic volume, high complexity in the identification of endocardial border and tiny labelled data. Due to advancements of artificial intelligence, we were able to address the challenges through a technique called data augmentation. three different deep learning models have been applied, the first model is UNet model, while the second model is Deeplab, and the third model is UNet model with backbone network for automated segmentation of left ventricle. The deep learning model has successfully segmented the left ventricle region on apical 4-chamber (A4C) views. The first model of UNet has mean IOU of 74.31%, and the second model of Deeplab has mean IOU of 79.53%. The proposed model has achieved promising mean IOU of 89.1% in segmentation and predicted ejection fraction with a correlation coefficient r2 of 0.71. Results show that the contactless echocardiographic approach can quantitatively estimate left ventricular chamber size and ejection fraction in humans as well as estimate the dimensions of the left ventricle in systole and diastole.