Aprendizagem profunda para o suporte ao diagnóstico da pneumonia causada por Covid-19 em exames de raio x e tomografia computadorizada

Abstract

The exponential spread of COVID-19 around the world has brought important challenges to public health systems. We learned that in pandemic situations, the high volume of patients overwhelms the care capacity of primary health care services. It became clear that in this COVID-19 pandemic, to control the morbidity and mortality of the disease, it is necessary to quickly identify as many patients as possible with suspected pneumonia. Imaging tests have been successfully used to identify and confirm suspected pneumonia associated with COVID-19. COVID-19 patients usually have abnormal situations on chest images obtained by computed tomography (CT) and X-ray exams. Although the radiological findings are similar to those found in other lung diseases, specialists are able to identify images of ground-glass features and suggest the diagnostic possibility of COVID-19 pneumonia even in its initial phase, especially in pandemic or epidemic situations. In this context, this work presents the use of a deep learning algorithm on the images associated with the diagnosis of COVID-19. Two open source systems were developed, with convolutional neural networks, capable of performing the identification of images suggestive of COVID-19, present in chest´s X-ray and CT scans. For the training of neural networks, data were collected from international public databases as well as data obtained through partnerships with Santa Izabel hospitals in Salvador, Bahia, Hospital das Clínicas, in São Paulo, capital, and Medsenior hospital, in Vitória, Espírito Santo. For the development of the X-ray image classification algorithm, 44,031 exams were used during training and validation. The X-ray model obtained a sensitivity of 0.85, specificity of 0.82 and ROC AUC of 0.93 when tested on a set of 1,158 chest radiographs from a referral hospital. The computed tomography classification algorithm used a base of 3,000 exams and innovatively selected the 16 most representative images per exam for training. The algorithm achieved a sensitivity of 0.89, specificity of 0.90, ROC AUC of 0.97 over a test base of 414 samples. The two algorithms presented themselves as good options for evaluating the chest images of COVID-19 patients, making it possible to separate those with alterations from those with no abnormalities. This is relevant in supporting the targeting of suspected patients in unassisted regions that lacks more efficient diagnostic models. We conclude that the use of intelligent algorithms can help in the identification of abnormal images in X-ray and chest tomography, especially in the absence of a specialist to report the exam, allowing a quick triage of those patients with suspected pneumonia who need immediate attention in the healthcare service.