Tomographic Identification and Evaluation of Pulmonary Involvement Due to SARS-CoV-2 Infection Using Artificial Intelligence and Image Segmentation Technique

Abstract

In 2019 coronavirus pandemic was required to find ways to further early diagnosis of COVID-19. This research aims to implement deep neural networks. Thus, we sought work to present analyses about the Deep Learning model, CNN, COVID-19 detection, and segmentation image. The training dataset to train the model we used exams the computerized tomography (CT scan) with dataset the Harvard Dataverse, available since May 2020, was used. The data were collected from the Public Hospital of the Government of the Employees of São Paulo (HSPM) and the Hospital Metropolitano da Lapa - São Paulo - Brazil. We applied it to four different architectures like VGG 19, Resnet 50, Inception, Xception. Each architecture has its advantages and disadvantages, and to supply the needs that each architecture presents, we generated ensembles among them. In general, the segmentation has shown that it is possible to capture regions with COVID-19 and differentiate them from other diseases. This study has pointed out the accuracy of 95.05% with a low false-positive rate for detection using computed tomography imaging. Thus, for automatic image renewal to show lung involvement, preliminary results translate into the lung and area affected by SARS-CoV-2. Future works can be done to improve our results, in particular, more databases may include detection of multiple disease cases such as pneumonia, bronchitis. Thus, it may be indicated more information to the classification result.

Federal University of Alfenas-Minas Gerais-Brazil.