Detection of COVID-19 Using a Multi-scale Deep Learning Network: Covid-MSNet

Abstract

Currently, the novel coronavirus disease-19 (COVID-19) has become a worldwide serious public health problem. Screening methods based on chest radiological imaging (e.g., X-ray and computed tomography) have been widely used as a primary tool for early diagnosis and treatment of COVID-19. Many studies have used single-scale deep learning models for COVID-19 detection from chest radiological images using global features. In chest radiological images, the identification of regions of interest using single scale may be difficult because the disease often appears in different positions at varying scales. For a more accurate and early detection of COVID-19, it is necessary to consider local spatial (location attention) features along with the global features. In this paper, we propose such a deep learning network referred to as Covid-MSNet for early detection of COVID-19 based on multi-scale learning that considers both attention and global features. Covid-MSNet comprises three main layers, namely a convolutional layer, a multi-scale stream layer and a fusion layer. In the convolutional layer, convolutional and max pooling layers are integrated together. In the multi-scale stream layer, features are extracted from the convolutional layer output images on multiple scales. To select the most discriminant features, a saliency-based learning is employed in the fusion layer. Performance evaluation of the Covid-MSNet is presented using the publicly available COVIDx dataset consisting of chest X-ray images of normal, pneumonia and COVID-19 patients. The results are compared with that of other well-known deep learning models, namely Inception-V4, Xception and ResNet50, and the results indicate that an improvement in overall accuracy of 3–14% is achieved along with an improvement in specificity of 3–10% using Covid-MSNet. Additionally, the critical regions in the lungs that are getting affected due to COVID-19 are highlighted using the heat maps generated from the discriminating features obtained using Covid-MSNet.KeywordsCOVID-19Medical image classificationMulti-scale learningPneumonia