Deep Learning for COVID-19 Marker Classification and Localization in Point-of-care Lung Ultrasonography

Abstract

In light of the current COVID19 epidemic, some studies have begun to look into DL-based solutions for the aided detection of lung disorders. Deep learning (DL) has demonstrated effectiveness in medical imaging. This research investigates the use of DL approaches for the analysis of lung ultrasonography (LUS) images, whereas previous efforts have concentrated on CT scans. With labels identifying the severity of the illness at a frame-level, video-level, and pixel-level, we provide a brand-new fully-annotated collection of LUS pictures acquired from multiple Italian hospitals (segmentation masks). We offer a number of deep models using this data that address pertinent problems for the autonomous processing of LUS pictures. We introduce a brand-new deep network in particular, formed from spatial transformer networks, which, in a weakly-supervised manner, both locates pathological artefacts and predicts the illness severity score related to an input frame. We also provide a novel approach for efficient frame score aggregation at the video-level based on uniforms. Finally, we evaluate cutting-edge deep models for calculating COVID-19 imaging biomarker pixel-level segmentations. Research on DL for the aided diagnosis of COVID-19 using LUS data is now possible thanks to experiments on the suggested dataset demonstrating good results on all the tasks taken into consideration.