Uncertainty to Improve the Automatic Measurement of Left Ventricular Ejection Fraction in 2D Echocardiography Using CNN-Based Segmentation

The echocardiographic measurement of the left ventricular ejection fraction (LVEF) is the accepted clinical way to assess the cardiac function of a patient. For this measurement, a physician needs to identify the end-systole and end-diastole, segment the left ventricle in those frames, obtain the volume from the masks, and compute the LVEF. Naive implementations of convolutional neural network (CNN) based segmentation algorithms to perform this measurement might encounter problems identifying the end-systole and end-diastole if there is a single poorly segmented frame in the whole echocardiogram, which would ruin the measurement of LVEF and require manual review by a human operator. In this research article, we present how to use different uncertainty metrics to identify poorly segmented frames and quantify how these techniques improve the concordance between algorithm and human operator measurements in a population-based cohort of echocardiographic examinations.