Avaliação de redes neurais convolucionais com transferência de aprendizado para classificação de imagens de edema macular diabético

• Ano de defesa: 2024
• Autor(a) principal: Conto Junior, Gilberto Luis de
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Tecnológica Federal do Paraná; Ponta Grossa; Brasil; Programa de Pós-Graduação em Ciência da Computação; UTFPR
• Palavras-chave em Português: Tomografia de coerência óptica; Processamento de imagens - Técnicas digitais; Redes neurais (Computação); Transferência de aprendizagem; Diabetes; Edema; Macula lutea; Optical coherence tomography; Image processing - Digital techniques; Neural networks (Computer science); Transfer of training; CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO; Engenharia/Tecnologia/Gestão
• Link de acesso: http://repositorio.utfpr.edu.br/jspui/handle/1/34497

Diabetes Mellitus is a disease that affects approximately 422 million people around the world. Diabetic Retinopathy is a condition that impacts the small vessels of the retina, being one of the most common complications of diabetes, which can progress to blindness. The disease can affect the central area of the retina, the macula, causing diabetic macular edema. This edema can be diagnosed through an Optical Coherence Tomography (OCT) exam to observe the layers of the retina and identify signs of the disease, such as fluid leakage in the area. This work aims to use pre­trained models of convolutional neural networks, applying transfer learning for the classification of OCT images, and to perform a comparative analysis of these networks. The network architectures used were ResNet­50, Inception V3, AlexNet, GoogLeNet, and VGG­19. For the training of the networks, 22,696 OCT images were used, equally divided between normal exams and those with diabetic macular edema. The models were tested with 250 images of each class as previously defined in the dataset. After each model was trained and applied to the test base, it was observed that the GoogLeNet network presented the best performance in classification. The results obtained with GoogLeNet were an accuracy of 99.20%, a precision of 98.43%, a recall of 100%, an F­measure of 99.21%, and a specificity of 98.40%.