Identificação e quantificação automática de taxa de glomérulos hialinizados utilizando deep learning

• Ano de defesa: 2020
• Autor(a) principal: Costa, Thalita Munique
• 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: Glomérulos renais - Detecção; Aprendizado de máquinas; Processamento de imagens - Técnicas digitais; Histopatologia; Redes neurais (Computação); Simulação (Computadores); Kidney glomerulus - Detection; Machine learning; Image processing - Digital techniques; Histology, Pathological; Neural networks (Computer science); Computer simulation; CNPQ::ENGENHARIAS::ENGENHARIA BIOMEDICA; Engenharia Elétrica
• Link de acesso: http://repositorio.utfpr.edu.br/jspui/handle/1/28827

In Digital Pathology, histological slides are scanned for further analysis. Digitized slides allow the use of artificial intelligence and image processing techniques for automatic identification and quantification in histopathology, allowing the quantification of the rate of hyalinized glomeruli. In this work, a database with images from several renal pathology study centers is utilized and the usage of Deep Learning, specifically the YOLOV3 architecture, is evaluated in the automatic detection of glomeruli. In addition to the assessment of functional glomeruli, there is also a need to identify the presence and percentage of hyalinized glomeruli (i.e. glomeruli that have become non-functional due to the replacement of all histopathological elements with hyaline protein material), when considering the totality of existing glomeruli. Using the Bio Atlas database (Pennsylvania State University), 16 complete slides, which resulted in 1177 images of 1024x1024 pixels with one or more glomeruli, were used for training and validation. A total of 468 hyalinized glomeruli and 1261 non-hyalinized glomeruli were noted. Using the 53-layer convolutional neural network and input images adjusted to 512x512 pixels, this work obtained a sensitivity of 90%, precision of 96.9%, accuracy of 87.5% and an F1 score of 93.3% considering the two types of glomeruli. A system was created to identify functional and hyalinized glomeruli, allowing support for histopathological study of kidney diseases and facilitating the location of the objects of analysis.