Automatic estimation of the ejection fraction from diastole and systole ultrasound images by a simplified end-to-end U-Net neural network

Left ventricular ejection fraction (LVEF) is the most useful cardiac index to assess the cardiac function of patients from echocardiograms (ECHO). Cardiologists manually delineate the left ventricular (LV) contour at the end of the diastolic (EDV) and systolic (ESV) times to calculate LVEF. This paper presents a novel end-to-end deep-learning model to directly estimate LVEF from only a pairwise sample (EDV and ESV) of ECHO images, which consists of two stages. Firstly, a simplified U-Net is constructed by reducing the number of convolutional filters per layer and removing the deepest layer. Secondly, the encoding part of the U-Net is used to extract features from the input ECHO images and their differences are concatenated into a vector entered into interconnected convolutional layers that reduce the dimensionality and produce an output that has a simple linear regressor applied to it for obtaining the ejection fraction value. The experimental results show a mean square error (MAE) of 5.61 and a coefficient of determination R
²
of 0.76 of automatic estimation of LVEF. This suggests that a simplified end-to-end deep-learning approach is able to both directly estimate LVEF and auto-delineate the LV from only a pairwise sample (EDV and ESV) from ECHO, which could be more viable and efficient for clinical practice routine.