Spatial Frequency Analysis by DWT of CXR in the COVIDGR Dataset

Abstract

In this study, chest X-ray images from the “COVIDGR Dataset” were formalized, and both normal images and disease severity images (mild, moderate, and severe) were extracted by the discrete wave transform (DWT) method. The characteristics of Approximation, Horizontal, Vertical, and Diagonal were analyzed as normal, mild, and vector frequencies. The standardized pixel was a format of 512 x 512 images and was intended to include all lung fields, except the original image, did not include the apex of the lung. The standardized images were extracted by DWT for the characteristics of Approximation, Horizontal, Vertical, and Diagonal and stored as COVID-19 Feature Data. These data were entered into an Excel file to calculate the sum, and the average value was obtained and confirmed as the “Feature value of images by disease.” Experiments showed that images of severe disease had little or no Approximation, high Diagonal content, and similarity between the Horizontal and Vertical characteristics. Normal images had the largest the proportion of Approximation because there was no change in frequency due to the absence of disease, while mild and moderate images were similar in Approximation, Horizontal, and Vertical but had a one-sided bias in small areas for Diagonal. It is hoped that the results of this study can be used as an important parameter in an automatic identification system of chest X-ray images to allow timely diagnosis of patients with COVID-19 and other lung conditions.