Background
Computer vision and deep learning algorithms have been successfully used to aid interpretation of echocardiograms, automate view classification and estimate standard cardiac measurements. We developed a deep learning model to detect pericardial effusions, which can be readily used for point-of-care ultrasound.
Methods
Using a dataset of 800 echocardiograms from 2016 to 2018, we trained a convolutional neural network to detect the presence of pericardial effusion from apical four chamber and subcostal view still images. Approximately half of the echocardiograms contained effusion. The network was based on Inception-v3, a deep convolutional network with residual connections, size invariant convolutions and modular design. It was built using Python and TensorFlow. The dataset was split by patient into 27,000/3,000/7,000 images for model training, validation and testing.
Results
Our convolutional neural network successfully detected pericardial effusion in apical four chamber view images with an accuracy of 91%(AUC 0.95) and in subcostal view images with an accuracy of 87%(AUC 0.94). In a sample of 51% positive apical four chamber images, the positive predictive value was 95% and the negative predictive value was 91%. In a sample of 49% positive subcostal images, the positive predictive value was 88% and the negative predictive value was 84%.