Uma metodologia para a criação de arquiteturas de redes neurais convolucionais otimizadas para a detecção e o diagnóstico entre pneumonias

Abstract

Pneumonia is a disease that affects the lungs, making it difficult to breathe. According to published studies, pneumonia is the disease that kills most children under the age of five, with most deaths occurring in developing countries. With the emergence of COVID-19, pneumonia has again become a worldwide concern. A great effort by companies, governments, the medical and academic community has been made worldwide to contain the spread of the coronavirus. The rapid spread of the SARS-CoV-2 virus makes early diagnosis extremely important. Thus, automation or acceleration in the diagnostic process is desirable. The use of computational methods can decrease specialists’ workload and even offer a second opinion, increasing accurate diagnoses. This thesis proposes a methodology for constructing convolutional neural network architectures specific for pneumonia detection and classification between bacterial and viral types (including COVID-19) through Bayesian optimization of pre-trained networks. The results obtained are promising, reaching, among other metrics, 100% accuracy for COVID-19 detection and more than 95% accuracy in the diagnosis between the normal, bacterial, viral, and COVID-19 types.