A Super-resolution-based Approach for the Detection of Covid—19 Infection From Chest X-ray Images

Abstract

X-ray is the most accessible imaging modality for detecting Covid-19 infection. However, X-ray image resolution depends on the amount of radiation dose. The Lesser the dosage, the lower the resolution, the higher the noise and patient safety. Detecting Covid-19 infection would be more precise with high-resolution chest X-ray images. The current article explores an edge-preserving single-scale residual learning-based super-resolution method to enhance low-resolution chest X-ray images. We used unsharp masking to preserve small, medium, and high-scale details while super-resolving the given image. The method produces a clear view of the pulmonary opacities in chest X-ray images after super-resolution reconstruction. Statistical feature metrics of first and second-order showed superior quality reconstruction by the proposed method for the given Covid-19 chest X-ray images. Further, to measure the effectiveness of super-resolution, we used an Inception v3 based deep learning model to classify chest X-ray images of Covid-19, pneumonia, and normal class. The performance of the classification model with super-resolved chest X-ray images was tested against 400 images belonging to two different classes at a time. We obtained increased precision of 94% and 96% accuracy in detecting Covid-19 infection in chest X-ray images after super-resolution compared to 64% precision and 68% accuracy before super-resolution.