Análise do uso de redes neurais com algoritmos evolutivos para a classificação de imagens histológicas de câncer de mama
Arquivos
Data
2022-08-04
Autores
Miguel, João Pedro Miranda 
Orientadores
Tosta, Thaína Aparecida Azevedo
Tipo
Trabalho de conclusão de curso
Título da Revista
ISSN da Revista
Título de Volume
Resumo
Tendo em vista que o câncer de mama é um dos tipos de câncer mais incidentes no
Brasil e que os exames de biópsia, utilizados na identificação e confirmação da doença,
são demorados e complexos, neste trabalho foi analisado o uso de algoritmos evolutivos
no treinamento de redes neurais para a classificação de imagens histológicas coradas por
hematoxilina-eosina a fim de auxiliar nesse processo. Para isso, foi analisada a aplicação dos
algoritmos de evolução diferencial, evolução diferencial com ilhas e migração, otimização
por enxame de partículas, otimização adaptativa por enxame de partículas e o algoritmo
genético. A classificação das imagens proposta utilizou atributos de cor do modelo de cores
RGB, atributos de textura por padrões binários locais e de redes neurais convolucionais
pela arquitetura Resnet50. Os parâmetros dos algoritmos evolutivos foram otimizados
utilizando o método de configuração de algoritmos baseado em modelo sequencial. Os
resultados obtidos foram comparados com trabalhos da literatura, bem como com a rede
neural artificial por retropropagação e com outros algoritmos clássicos, como a máquina
de vetores de suporte e as florestas de decisão aleatória. Os algoritmos de otimização por
enxame de partículas e otimização adaptativa por enxame de partículas foram os que
apresentaram os resultados mais promissores, com o primeiro alcançando AUC máxima
de 0,84 para atributos de textura, e o último com AUC máxima também de 0,84 para os
atributos de cor, ambos na base DCIS-UDH.
Bearing in mind that breast cancer is one of the most common types of cancer in Brazil and that biopsy tests, used in the identification and confirmation of the disease, are time-consuming and complex, the use of evolutionary algorithms in the training of neural networks was analyzed in this work for the classification of histological images stained by hematoxylin-eosin to assist in this process. For this purpose, the algorithms differential evolution, differential evolution with islands and migration, particle swarm optimization, adaptive particle swarm optimization, and genetic algorithm are analyzed, considering color features of the RGB color model, texture features of local binary patterns, and convolutional neural networks through the Resnet50 architecture. The parameters of the evolutionary algorithms were optimized using the sequential model-based algorithm configuration method. The obtained results were compared with studies of the literature, as well as with the backpropagation artificial neural networks and with other classic algorithms such as the support vector machine and the random forest. The particle swarm optimization and adaptive particle swarm optimization presented the most promising results, with the first one reaching a maximum AUC of 0.84 for the texture features, and the last one also with a maximum AUC of 0.84 for the color attributes, both on DCIS-UDH dataset.
Bearing in mind that breast cancer is one of the most common types of cancer in Brazil and that biopsy tests, used in the identification and confirmation of the disease, are time-consuming and complex, the use of evolutionary algorithms in the training of neural networks was analyzed in this work for the classification of histological images stained by hematoxylin-eosin to assist in this process. For this purpose, the algorithms differential evolution, differential evolution with islands and migration, particle swarm optimization, adaptive particle swarm optimization, and genetic algorithm are analyzed, considering color features of the RGB color model, texture features of local binary patterns, and convolutional neural networks through the Resnet50 architecture. The parameters of the evolutionary algorithms were optimized using the sequential model-based algorithm configuration method. The obtained results were compared with studies of the literature, as well as with the backpropagation artificial neural networks and with other classic algorithms such as the support vector machine and the random forest. The particle swarm optimization and adaptive particle swarm optimization presented the most promising results, with the first one reaching a maximum AUC of 0.84 for the texture features, and the last one also with a maximum AUC of 0.84 for the color attributes, both on DCIS-UDH dataset.