Abstract
Background: CT scans give highly detailed lung images, but patient transport, machine preparation, and scanning time are limiting factors. Chest x-rays are more widely available, but images are non-specific and often difficult to interpret.
Aims and Objectives: To develop a machine-learned early detection system, using conventional chest x-rays, to identify COVID-19 and differentiate its phenotypes.
Methods: We extracted features of chest x-rays using a DenseNet-121 neural network. We then sorted the images according to similarity using the UMAP clustering algorithm (Uniform Manifold Approximation and Projection). Both algorithms were unsupervised, in the sense that data labels for presence or absence of disease were known but only provided afterward for visualization; they were not seen or used by the algorithms.
Results: Our method achieved a 95.19% accuracy in the COVIDx dataset (Wang et al. arXiv 2020). Its performance for detecting normal, pneumonia, and COVID-19 cases is summarized in Table 1. UMAP clustering for the x-ray images is shown in Fig. 1. Location in the UMAP space gives a quantitative severity scale for COVID-19.
Conclusions: Feature extraction can identify unique signatures of COVID-19 chest x-rays. A nonlinear clustering algorithm reveals two types of COVID-19 response: one resembling pneumonia and one with a more normal presentation.
Footnotes
Cite this article as: European Respiratory Journal 2020; 56: Suppl. 64, 4151.
This abstract was presented at the 2020 ERS International Congress, in session “Respiratory viruses in the "pre COVID-19" era”.
This is an ERS International Congress abstract. No full-text version is available. Further material to accompany this abstract may be available at www.ers-education.org (ERS member access only).
- Copyright ©the authors 2020