Análise de histopatologia renal usando deep learning

• Ano de defesa: 2020
• Autor(a) principal: Barbosa, Lourenço Madruga
• 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 Engenharia Elétrica e Informática Industrial; UTFPR
• Palavras-chave em Português: Aprendizado de máquinas; Redes neurais (Computação); Glomérulos renais; Histopatologia; Processamento de imagens - Técnicas digitais; Machine learning; Neural networks (Computer science); Kidney glomerulus; Histology, Pathological; Image processing - Digital techniques; CNPQ::ENGENHARIAS::ENGENHARIA BIOMEDICA; Engenharia Elétrica
• Link de acesso: http://repositorio.utfpr.edu.br/jspui/handle/1/27820

In this work we use convolutional neural network and deep learning to detect objects of biomedical interest (i.e., glomeruli) in whole digitalized slides from studies of renal pathology. The analysis of glomeruli is important for studies of nephropathies caused by diabetes, lupus, drug use, excess of drugs, among other causes. The digitization of entire slides allows the application of a variety of digital image processing techniques and the use of artificial intelligence to assist doctors in diagnoses. We carried out the of a YOLOv3 convolutional neural network of architecture with 53 layers. The training and performance analysis of the network had 6 experiments. The training data set contained a total of 16 entire slides, divided into sub-images of 2048x2048, which resulted in a total of 815 images, containing a total of 2325 annotated glomeruli, for training, validation and performance evaluation. For the performance analysis of the trained network (i.e., detection), 7 unknown slides of the network were selected. Based on the result found, it is possible to state that the result of this work is superior to those found in the literature. Accuracy, Precision, Sensitivity, Specificity and F1s of 99.40%, 97.31%, 96.17%, 99.73% and 94.24%, respectively, were obtained in this work showing the potential to assist in the diagnosis of histopathological exams.