An automated method for segmentation of COVID-19 lesions based on Computed Tomography using deep learning methods

SARS-CoV-2, a disease caused by the new coron-avirus, after discovery, spread rapidly to several countries. The number of infected people exceeds 176 million worldwide, and more than 3.8 million deaths are recorded. There are several ways to diagnose COVID-19, ranging from rapid tests to imaging exams. This article proposes a computational methodology that combines two Deep Learning techniques, one based on the U-Net architecture and the other on Adversarial Generating Networks (GAN), with Computed Tomography (CT) to investigate suspicious regions of COVID-19. After the false positive reduction step with the application of morphological operations, the results achieved were promising. For the U-Net architecture we reached DICE 0.754, IoU 0.606, Sensitivity 0.825, Specificity 0.998, Accuracy 0.997, AUC 0.912, Accuracy 0.694 and F-Score 0.754. At GAN we achieved DICE 0.770, IoU 0.626, Sensitivity 0.747, Specificity 0.998, Accuracy 0.997, AUC 0.873, Accuracy 0.794 and F-Score 0.770.