An Image Segment-based Classification for Chest X-Ray Image

Abstract

In late 2019, the first case of COVID-19 was confirmed in Wuhan, China. The number of cases has been rapidly growing since then. Molecular and antigen testing methods are very accurate for the diagnosis of COVID-19. However, with sudden increases of infected cases, laboratory-based molecular test and COVID-19 test kits are in short supply. Because the virus affects an infected patient’s lung, interpreting images obtained from Computed Tomography Scanners and Chest X-ray Radiography (CXR) machines can be an alternative for diagnosis. However CXR interpretation requires experts and the number of experts is limited. Therefore, automatic detection of COVID-19 from CXR images is required. We describe a system for automatic detection of COVID-19 from CXR images. It first segmented images to select only the lung. The segmented part was then fed into a multiclass classification module, which worked well with samples obtained from various sources, which had different aspect ratios, contrast and viewpoints. The system also handled the unbalanced dataset—only a small fraction of images showed COVID-19. Our system achieved 92% of F1-score and 88.1% Marco F1-score on the 3rd Deep Learning and AI Summer/Winter School Hackathon Phase 3—Multi-class COVID-19 Chest X-ray challenge public leaderboard.