COVID-19 feature detection with deep neural networks trained on simulated lung ultrasound B-mode images

Deep learning has been implemented to detect COVID-19 features in lung ultrasound B-mode images. However, previous work primarily relied on in vivo images as the training data, which suffers from limited access to required manual labeling of thousands of training image examples. To avoid this manual labeling, which is tedious and time consuming, we propose the detection of in vivo COVID-19 features (i.e., A-line, B-line, consolidation) with deep neural networks (DNNs) trained on simulated B-mode images. The simulation-trained DNNs were tested on in vivo B-mode images from healthy subjects and COVID-19 patients. With data augmentation included during the training process, Dice similarity coefficients (DSCs) between ground truth and DNN predictions were maximized, producing mean ± standard deviatio values as high as 0.48 ± 0.29, 0.45 ± 0.25, and 0.46 ± 0.35 when segmenting in vivo A-line, B-line, and consolidation features, respectively. Results demonstrate that simulation-trained DNNs are a promising alternative to training with real patient data when segmenting in vivo COVID-19 features.