On the impact of pixel resolution on automated scoring of lung ultrasound images from coronavirus disease 2019 patients

Abstract

With the outbreak of the COVID-19 pandemic, remote diagnosis, patient monitoring, collection, and transmission of health data from electronic devices are rapidly taking its share in the health sector. These devices are, however, limited on resources like energy, memory, and processing power. Consequently, it is highly relevant to investigate how to minimize the size of data, keeping intact the information content. The objective of this study is to, thus, observe the impact of pixel resolution on the automated scoring by DL algorithms for LUS videos. First, 448 videos from 20 patients were normalized to a common pixel resolution, i.e., the largest found over the dataset (841 pixels/cm2). Next, the pixel resolution was further reduced by factor 2 by resampling the normalized data to 210 pixels/cm2. Original, normalized and re-sampled videos were evaluated using the DL algorithm [Roy et al., IEEE Trans. Med. Imaging 39, 2676–2687 (2020)]. At frame level, for normalized and resampled videos, the level of agreement of DL results with the original videos was 93.2% and 86.6%, respectively. Similar performance was found at video level with the agreement to 95.75% and 85.93%, respectively. The study showed that with a significant reduction in the pixel resolution of LUS data, low variation in the DL performance is observed.