Automatic image processing to identify post-COVID conditions by using deep learning

Abstract

In the present research, a supervised learning classification methodology is proposed to identify post-COVID conditions. Image processing and deep learning methods were employed to analyze a data set provided by the High Specialty Medical Unit No.1 of the Mexican Institute of Social Security (T1-IMSS) of Leon, Guanajuato, Mexico, of Mexican patients infected with COVID-19. The dataset is classified into post-COVID findings and no post-COVID findings. A deep neural network of 50 hidden layers is used to extract regions of interest, with properties that can potentially be related to computer-aided medical diagnosis. Different patterns were found in the post-COVID computed tomography scans: pulmonary fibrosis, ground glass pattern, etc. The efficiency of the proposed method was 97% precision using the cross-validation classification scenario. This result allows to provide an auxiliary tool in medical diagnosis, through computer-aided diagnosis. This model provides an automatic and objective estimation of post-COVID conditions of Mexican patients, facilitating the expert interpretation during the COVID-19 pandemic.