Development and validation of an end-to-end deep learning pipeline to measure pericardial effusion in echocardiography

IntroductionCardiac tamponade, caused by pericardial effusion (PE), is a life-threatening condition that can be resolved by timely pericardiocentesis. Nevertheless, PE measurement remains operator-dependent and may be difficult in some circumstances. Our study aimed to develop a deep-learning pipeline that measures the amount of PE based on raw echocardiography clips. MethodsEchocardiographic examination data were collected from one medical center in southern Taiwan from 2010-2018. Four commonly used cardiac windows, including the parasternal long-axis, parasternal short-axis, apical four-chamber, and subcostal views from included ultrasound examinations, were used for analysis. We proposed a deep learning pipeline consisting of three steps: moving window view selection, automated segmentation, and width calculation from a segmented mask. The pipeline was then prospectively validated from 2019-2020 using a dataset from the same hospital, and externally validated using data from another medical center in Taiwan. Model performance was evaluated using mean absolute error, intraclass correlation coefficient (ICC), and R-squared value between the ground truth and predictions. ResultsIn this study, 995 echocardiographic examinations were included. Among these, 155 were used for internal validation and 258 were used for external validation. The proposed pipeline had a predictive performance of ICC=0.867 for internal validation and ICC=0.801 for external validation. It accurately detected PE with an area under the receiving operating characteristic curve (AUC) of 0.926 (0.902-0.951) for internal validation and 0.842 (0.794-0.889) for external validation. Regarding the recognition of moderate PE or worse, the AUC values improved to 0.941 (0.923--0.960) and 0.907 (0.876-0.943) for internal and external validation, respectively. Of all the selected cardiac windows, our model had the best prediction in the parasternal long-axis and apical four-chamber views. ConclusionsThe machine-learning pipeline could automatically calculate the width of the PE from raw ultrasound clips. The novel concepts of moving window view selection for image quality control and computer vision techniques for maximal PE width calculation seem useful in the field of ultrasound.