Lung ultrasound patterns analysis at video and prognostic level in a resource-constrained setting

Low-cost and ultra-portable point-of-care ultrasound (POCUS) devices can now be used to aid in diagnosis and testing in limited resource settings. However, interpretation of data is one of the key barriers to POCUS adoption and implementation. Artificial intelligence can potentially address this issue, by reducing the impact of poor confidence in data interpretation. Interpreting lung ultrasound (LUS) mainly includes the analysis of hyper-echoic horizontal and vertical artifacts, and hypo-echoic small to large consolidations. Our aim is to design a computationally efficient method of interpretation and classification of LUS patterns in a video, effectively deployable for POCUS analysis in a resource-constrained environment. In this regard, we propose an intensity projection (IP)-based technique to compress LUS patterns within a video. The method works by projecting maximum, mean, and minimum intensity values across the temporal dimension of a video to form a compressed image. This allows to preserve hyper- and hypo-echoic patterns while compressing a video to three images. The compressed images are then classified using a convolutional neural network. Results show that IP-based compression effectively preserves LUS patterns in compressed images. Moreover, the classification of these images achieves state-of-the-art performance while reducing the number of frames needed to assess a LUS video from an average of few hundreds down to three images. This approach lays down the foundation for efficient and accurate video- and prognostic-level stratification in resource-constrained environments.