Emprego de Swin Transformer para classificar imagens radiográficas de tórax e diagnosticar COVID-19

• Ano de defesa: 2023
• Autor(a) principal: Ferraz, Aroldo
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Tecnológica Federal do Paraná; Curitiba; Brasil; Programa de Pós-Graduação em Computação Aplicada; UTFPR
• Palavras-chave em Português: Imagens - Radiografia; Visão por computador; COVID-19 (Doença) - Diagnóstico; Aprendizado profundo (Aprendizado do computador); Tomografia; Images - Radiography; Computer vision; COVID-19 (Disease) - Diagnosis; Deep learning (Machine learning); Tomography; CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO; CNPQ::ENGENHARIAS::ENGENHARIA BIOMEDICA; Ciência da Computação
• Link de acesso: http://repositorio.utfpr.edu.br/jspui/handle/1/32649

According to data from the World Health Organization (WHO), since the beginning of 2020, the COVID-19 pandemic has infected over 770.56 million people worldwide. Of those infected, more than 6.95 million individuals have lost their lives. COVID-19 is an extremely contagious disease and can swiftly overwhelm healthcare systems if infections are not diagnosed, and if isolation measures and treatment for deteriorating conditions are not promptly implemented. The primary screening test used to diagnose COVID-19 has been the Real Time Reverse Transcription-Polymerase Chain Reaction (RT-PCR). Although it is accurate and reliable, it takes a significant amount of time to deliver a definitive diagnosis. Since the onset of the pandemic, several researchers have demonstrated the feasibility of employing various Deep Learning (DL) techniques based on Convolutional Neural Networks (CNN), which have shown promising performance in diagnosing COVID-19 from X-rays and CT scans. However, some authors argue that existing DL methods, based on CNN, are being challenged by newer DL architectures and models. Motivated by this, in this study, we propose the use of the Swin Transformer (instead of CNN) for COVID-19 screening using X-ray images from four original datasets: COVID-QU-Ext Dataset, SARS-COV2-CT Dataset, HUST-19, and HCV-UFPR-COVID-19. These datasets were merged using different strategies, resulting in a total of 9 distinct datasets. We employed transfer learning to address the issue of data scarcity. Moreover, we assessed whether the models have a high generalization capability, based on the learned features, allowing training on one dataset, and testing on another, aiming to evaluate if there are significant losses in the metric values obtained. Upon analyzing the cumulative results from experiments across all datasets and strategies, the Swin Transformer stands out. In terms of sensitivity/recall, it performed 107% and 80% better than ViT and CNN, respectively. In the HUST-19 dataset, the Swin Transformer achieved the maximum score (1 or 100%) for all metrics and continued to show notable performance in other datasets. Compared to the state of the art, the results reveal a promising and highly competitive performance for the Swin Transformer. Additionally, statistical tests were conducted which showed, in several instances, that there are statistically significant differences in the metrics obtained by the Swin Transformer when compared to CNN and ViT.