COVID-19 Severity Classification Using a Hierarchical Classification Deep Learning Model

Abstract

One of the most important situations in recent years has been originated by the 2019 Coronavirus disease (COVID-19). Nowadays this disease continues to cause a large number of deaths and remains one of the main diseases in the world. In this disease is very important the early detection to avoid the spread, as well as to monitor the progress of the disease in patients, and techniques of artificial intelligence (AI) is very useful for this. This is where this work comes from, trying to contribute in the study to detect infected patients. Drawing inspiration from previous work, we studied the use of deep learning models to detect COVID-19 and classify the patients with this disease. The work was divided into three phases to detect, evaluate the percentage of infection and classify patients of COVID-19. The initial stage use CNN Densenet-161 models pre-trained to detects the COVID-19 using multi-class X-Ray images (COVID-19 vs. No-Findings vs. Pneumonia), obtaining 88.00% in accuracy, 91.3% in precision, 87.33% in recall, and 89.00% in F1-score. The next stage also use CNN Densenet-161 models pre-trained to evidenced the percentage of infection COVID-19 in the different CT-scans slices belonging to a patient, obtaining in the evaluation metrics a result of 0.95 in PC, 5.14 in MAE and 8.47 in RMSE. The last stage creates a database of histograms of different patients using their lung infections and classifies them into different degrees of severity using K-Means unsupervised learning algorithms with PCA.