Use este identificador para citar ou linkar para este item:
Título: Síntese de imagens de Raios-X de problemas respiratórios utilizando redes neurais adversariais generativas
Autor(es): Wyzykowski, André Brasil Vieira
Melo, Osvaldo Requião
Reis, Marcelo Índio dos
Palavras-chave: Modelos generativos adversariais
Raio-X torácico
Generative adversarial models
Chest X-ray
Data do documento: 22-Jul-2020
Editor: Universidade Católica do Salvador
Resumo: Diverses problems in scientific study development are related to information scarcity. In the medical context, there are rares diseases that have a few cases. Beyond that, in initial stages of a new disease, patient data keep stored in medical centers, until they realize a compilation work of them. Lastly, still there a situation of about personal data are kept safe under law protection, causing a data unbalancing. Currently, this difficulty was evidenced with Coronavirus disease 2019 (COVID-19). As it’s a new disease, the scientific community had little precise data to carry out more detailed studies. Thinking about this, the work goal is to conceive x-ray lung synthetic images, which were generated by artificial intelligence, being visually realistic from a few real samples to enable the creation of repositories that will support future research and detailed studies. To reach the objective, one research was carried out to find the next works to support the development, experiments, and validations that contributed to the stage of training neural network models. To validate the provided information of that training was necessary to calculate the quality related to the original samples, to do that was used the calculation of Fréchet distance, to measure the characteristics interval between them, contributing to the validation to the best neural model applied. Counting too with a human perception experiment was applied in 81 participants in visual form, to validate by his optics and judge between the presented images in your criteria to consider the sample as genuine or synthetic. With the result obtained from the human perception experiment, was possible to identify that the samples judged by the participants as the closest to reality were the same as from the score calculated by the distance from Fréchet. With these validations, it was possible to demonstrate that the images generated by the model Wasserstein generative adversarial network (WGAN) were superior to the Deep Convolutional Generative Adversarial Network (DCGAN).
Aparece nas coleções:Engenharia de Software

Arquivos associados a este item:
Arquivo Descrição TamanhoFormato 

Os itens no repositório estão protegidos por copyright, com todos os direitos reservados, salvo quando é indicado o contrário.