The Impact of Histogram Equalization and Color Mapping on ResNet-34's Overall Performance for COVID-19 Detection

Abstract

The COVID-19 pandemic has had a “devastating” impact on public health and well-being around the world. Early diagnosis is a crucial step to begin treatment and prevent more infections. In this sense, early screening approaches have demonstrated that in chest radiology images, patients present abnormalities that distinguish COVID-19 cases. Recent studies based on Convolutional Neural Networks (CNNs), using radiology imaging techniques, have been proposed to assist in the accurate detection of COVID-19. Radiology images are characterized by the opacity produced by “ground glass” which might hide powerful information for feature analysis. Therefore, this work presents a methodology to assess the overall performance of Resnet-34, a deep CNN architecture, for COVID-19 detection when pre-processing histogram equalization and color mapping are applied to chest X-ray images. Besides, to enrich the available images related to COVID-19 studies, data augmentation techniques were also carried out. Experimental results reach the highest precision and sensitivity when applying global histogram equalization and pink color mapping. This study provides a point-of-view based on accuracy metrics to choose pre-processing techniques that can improve CNNs performance for radiology image classification purposes.