Aprendizado profundo para auxiliar a detecção de COVID-19 baseado em imagens de raio-x de tórax: uma abordagem prática

Abstract

Given the large number of COVID-19 cases around the world, a practical solution to reduce and alleviate the patient queue in hospitals and healthcare systems is welcome. Fast and reliable diagnostics based on technological tools can help medical professionals to manage this bottleneck situation. In this work, we propose a practical methodology using deep learning to detect and classify lungs affected by COVID-19 through chest x-ray imaging. RetinaNet architecture is used in the process. This architecture is a onestage detector combined with Focal Loss. We considered a dataset with 5000 images, 2000 to train the model, 1000 to validate, and 2000 to test the model. The results obtained show a recall score of 0.99, precision of 0.99, sensibility of 0.56, and mAP of 0.81. The high recall score implicates that a patient with COVID-19 will be classified correctly.