A novel machine learning-based video classification approach to detect pneumonia in COVID-19 patients using lung ultrasound

Context: Efficiently diagnosing COVID-19-related pneumonia is of high clinical relevance. Point-of-care ultrasound allows detecting lung conditions via patterns of artifacts, such as clustered B-lines. Aims: The aim is to classify lung ultrasound videos into three categories: Normal (containing A-lines), interstitial abnormalities (B-lines), and confluent abnormalities (pleural effusion/consolidations) using a semi-automated approach. Settings and Design: This was a prospective observational study using 1530 videos in 300 patients presenting with clinical suspicion of COVID-19 pneumonia, where the data were collected and labeled by human experts versus machine learning. Subjects and Methods: Experts labeled each of the videos into one of the three categories. The labels were used to train a neural network to automatically perform the same classification. The proposed neural network uses a unique two-stream approach, one based on raw red-green-blue channel (RGB) input and the other consisting of velocity information. In this manner, both spatial and temporal ultrasound features can be captured. Statistical Analysis Used: A 5-fold cross-validation approach was utilized for the evaluation. Cohen's kappa and Gwet's AC1 metrics are calculated to measure the agreement with the human rater for the three categories. Cases are also divided into interstitial abnormalities (B-lines) and other (A-lines and confluent abnormalities) and precision-recall and receiver operating curve curves created. Results: This study demonstrated robustness in determining interstitial abnormalities, with a high F1 score of 0.86. For the human rater agreement for interstitial abnormalities versus the rest, the proposed method obtained a Gwet's AC1 metric of 0.88. Conclusions: The study demonstrates the use of a deep learning approach to classify artifacts contained in lung ultrasound videos in a robust manner.