Desenvolvimento de um Sistema de Auxílio ao Diagnóstico de Pneumonia na Infância Utilizando Visão Computacional

• Ano de defesa: 2012
• Autor(a) principal: MACEDO, Sanderson Oliveira de
• Orientador(a): OLIVEIRA, Leandro Luís Galdino de
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Goiás
• Programa de Pós-Graduação: Mestrado em Ciência da Computação
• Departamento: Ciências Exatas e da Terra - Ciências da Computação
• País: BR
• Palavras-chave em Português: CBIR; Pneumonia; Textura; Wavelet; radiografia
• Palavras-chave em Inglês: CBIR; Pneumonia; Texture; Wavelet; Radiography
• Área do conhecimento CNPq: CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO
• Link de acesso: http://repositorio.bc.ufg.br/tede/handle/tde/519

This work aims to develop a system to aid in the diagnosis of pneumonia by computer, termed pneumocad, which aims to identify chest radiographies compatible with the disease. Techniques were used for the recognition of patterns in textures through the decomposition of the wavelet transforms of the features extracted from the decomposition and classification applied to radiography. We analyzed 166 images in digital radiography "gold standard", previously confirmed by two radiologists trained according to WHO guidelines as Pneumonia Present (PP = 83) and Pneumonia Absent (PA = 83). In both methods were investigated which feature best applies to the recognition of patterns and textures in which the best performance of the classifier K-NN method. The procedure began with the application of the Haar Wavelet Transform and the extraction of characteristics of each radiograph 17 that were stored descriptors. The methodology I tested the increase in classification accuracy, balanced with increasing the amount of radiographies of each class. The methodology II tested the ability of K-NN to generate ratings at acceptable levels with the unbalance of the random number of images between the two classes. There was obtained an average of accuracy of 91.75% with emphasis on the difference of characteristic variance performance and the K-NN was more effective when the number of nearest neighbors is K=9. The results are considered promising because the pneumocad can be a useful tool in the diagnosis of childhood pneumonia, combining the knowledge manmachine and providing conditions for the interpretation of chest radiographs in the "gold standard", according to WHO specifications. In addition, the software can be a new technology in health, to provide health managers and policymakers a tool for epidemiological monitoring and control of pneumonia in real time, producing benefits for organizations and decisions related health services.