Quantifying lung ultrasound comets with a convolutional neural network: Initial clinical results

Abstract

Lung ultrasound comets are "comet-tail" artifacts appearing in lung ultrasound images. They are particularly useful in detecting several lung pathologies and may indicate the amount of extravascular lung water. However, the comets are not always well defined and large variations in the counting results exist between observers. This study uses a convolutional neural network to quantify these lung ultrasound comets on a 4864-image clinical lung ultrasound dataset labeled by the authors. The neural network counted the number of comets correctly on 43.4% of the images and has an intraclass correlation (ICC) of 0.791 with respect to human counting on the test set. The ICC level indicates a higher correlation level than previously reported ICC between human observers. The neural network was then deployed and applied to a clinical 6272-image dataset. The correlation between the automated comet counts and the clinical parameters was examined. The comet counts correlate positively with the diastolic blood pressure (p = 0.047, r = 0.448), negatively with ejection fraction (p = 0.061, r = -0.513), and negatively with BMI (p = 0.009, r = -0.566). The neural network can be alternatively formulated as a diagnostic test for comet-positive images with 80.8% accuracy. The results could potentially be improved with a larger dataset and a refined approach to the neural networks used.

Keywords: B-Lines; Lung; Lung comets; Machine learning; Neural network; Ultrasound.