Utilização de redes neurais completamente convolucionais para identificação e medição de crânios fetais

• Ano de defesa: 2019
• Autor(a) principal: Skeika, Everton Leonardo
• 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: Ultrassonografia; Craniometria; Aprendizagem; Redes neurais (Computação); Ultrasonic imaging; Craniometry; Learning; Neural networks (Computer science); CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO; Engenharia/Tecnologia/Gestão
• Link de acesso: http://repositorio.utfpr.edu.br/jspui/handle/1/4714

Ultrasonography is a non-invasive, non-radioactive diagnostic imaging technique often used to monitor fetal development during pregnancy. An accurate assessment is important for the analysis of fetal growth, malformations diagnosis and possible congenital diseases, thus ensuring the well-being of the mother and fetus during pregnancy. However, accurate measurement of fetal anatomical structures requires specialized knowledge of the obstetrician. In addition to being a tedious and timeconsuming process, contour extraction is influenced by his experience. Given this issue in mid-2018, the HC18 challenge was launched to challenge researchers in the field to develop an algorithm that can automatically measure head circumference using two-dimensional ultrasound images. In this sense, this work proposes a solution to the challenge consisting of an adaptation of the Fully Convolutional Neural Network V-Net. To serve this objective, a methodology was built consisting of some steps, such as performing the pre-processing in the dataset provided by the challenge, changes in the network structure for better segmentation of images, the construction of an algorithm for post-processing of inefficient segmentation, the implementation of an algorithm for calculating the ellipses resulting from segmentation. Subsequently, the results obtained by the proposed methodology were submitted to the challenge. According to the classification obtained, it is shown that it is possible to use the V-Net network for this type of problem.