Ultrasound-Based Detection of Lung Abnormalities Using Single Shot Detection Convolutional Neural Networks

Ultrasound imaging can be used to identify a variety of lung pathologies, including pneumonia, pneumothorax, pleural effusion, and acute respiratory distress syndrome (ARDS). Ultrasound lung images of sufficient quality are relatively easy to acquire, but can be difficult to interpret as the relevant features are mostly non-structural and require expert interpretation. In this work, we developed a convolutional neural network (CNN) algorithm to identify five key lung features linked to pathological lung conditions: B-lines, merged B-lines, lack of lung sliding, consolidation and pleural effusion. The algorithm was trained using short ultrasound videos of in vivo swine models with carefully controlled lung conditions. Key lung features were annotated by expert radiologists and snonographers. Pneumothorax (absence of lung sliding) was detected with an Inception V3 CNN using simulated M-mode images. A single shot detection (SSD) framework was used to detect the remaining features. Our results indicate that deep learning algorithms can successfully detect lung abnormalities in ultrasound imagery. Computer-assisted ultrasound interpretation can place expert-level diagnostic accuracy in the hands of low-resource health care providers.