Deep learning-based pneumothorax detection in ultrasound videos

Abstract

Pneumothorax (PTX) is a medical and surgical emergency that can lead to hemodynamic instability and life-threatening collapse of the lung. PTX is usually detected using chest X-ray but can be detected using lung ultrasound, which requires interpretation by an expert radiologist. We are developing an AI based algorithm for the automated interpretation of lung ultrasound video to enable fast diagnosis of pneumothorax at the point of care by health care providers without extensive training in ultrasound. In this work, we developed and compared several deep learning methods for identifying pneumothoraces in 3-s ultrasound videos collected with a handheld ultrasound system. The first group of methods were based on convolutional neural networks (CNNs) paired with time-mapping preprocessing algorithms, including reconstructed M-mode and the proposed simplified optical flow transform (SOFT). These preprocessing methods were either used alone or in combination in a single “fusion” CNN. The second class of algorithm used a Deep Learning architecture that combines a CNN for processing spatial information (Inception V3) with a recurrent network (long-short-term-memory, or LSTM) for temporal analysis, enabling raw video to be fed directly into the neural network without preprocessing. We used data from a swine pneumothorax model to train and test the proposed algorithms, comparing their performance. Despite limited data, all algorithms achieved an AUC for pneumothorax detection greater than 0.83.