Deep Learning-Based CAD System for COVID-19 Diagnosis via Spectral-Temporal Images

Abstract

The diagnosis of COVID-19 and understanding the condition of the patients who have critical responses is crucial to stop the rapid propagation of such disease. Consequently, diminishing adverse impacts that affected various industrial divisions, especially healthcare. Deep learning methods have proven their great capabilities in studying and analyzing computed tomography (CT) images containing COVID-19. Most related studies utilized the spatial information of CT images to train deep learning models. Nevertheless, training these models with spatial-temporal images could enhance diagnostic accuracy. This paper proposes a computer-assisted diagnostic (CAD) system for COVID-19 diagnosis using three deep learning models trained with spectral-temporal images. First, it uses the multilevel discrete wavelet transform (DWT) to analyze the original CT images and obtain the spectral-temporal images. Then, it uses these images from different DWT levels to train three ResNets deep learning models. Afterward, for each ResNet trained with images of each DWT level, it extracts deep features. Next, for each ResNet, it fuses these deep features and then uses a feature selection approach to reduce their dimension. Finally, support vector machine (SVM) classifiers are used to perform classification. The performance of the proposed CAD proves that training ResNets with spectral-temporal images is better than using CT images. Also, the fusion and feature selection steps have enhanced the diagnostic accuracy, thus the proposed CAD could be employed to help radiologists in COVID-19 inspection.