Optimized chest X-ray image semantic segmentation networks for COVID-19 early detection

Abstract

BACKGROUND: Although detection of COVID-19 from chest X-ray radiography (CXR) images is faster than PCR sputum testing, the accuracy of detecting COVID-19 from CXR images is lacking in the existing deep learning models. OBJECTIVE: This study aims to classify COVID-19 and normal patients from CXR images using semantic segmentation networks for detecting and labeling COVID-19 infected lung lobes in CXR images. METHODS: For semantically segmenting infected lung lobes in CXR images for COVID-19 early detection, three structurally different deep learning (DL) networks such as SegNet, U-Net and hybrid CNN with SegNet plus U-Net, are proposed and investigated. Further, the optimized CXR image semantic segmentation networks such as GWO SegNet, GWO U-Net, and GWO hybrid CNN are developed with the grey wolf optimization (GWO) algorithm. The proposed DL networks are trained, tested, and validated without and with optimization on the openly available dataset that contains 2,572 COVID-19 CXR images including 2,174 training images and 398 testing images. The DL networks and their GWO optimized networks are also compared with other state-of-the-art models used to detect COVID-19 CXR images. RESULTS: All optimized CXR image semantic segmentation networks for COVID-19 image detection developed in this study achieved detection accuracy higher than 92%. The result shows the superiority of optimized SegNet in segmenting COVID-19 infected lung lobes and classifying with an accuracy of 98.08% compared to optimized U-Net and hybrid CNN. CONCLUSION: The optimized DL networks has potential to be utilised to more objectively and accurately identify COVID-19 disease using semantic segmentation of COVID-19 CXR images of the lungs.