Algoritmos bioinspirados para otimização de CNNs aplicados na detecção de câncer de mama utilizando imagens termográficas

• Ano de defesa: 2022
• Autor(a) principal: Gonçalves, Caroline Barcelos
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Uberlândia; Brasil; Programa de Pós-graduação em Ciência da Computação
• Palavras-chave em Português: Termografia; Câncer de mama; CNNs; AG; PSO; Surrogate; Random forest; Thermography; Breast cancer; Computação; CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO; Ciência da Computação; Mamas - Câncer - Prevenção; Mamas - Imagem
• Link de acesso: https://repositorio.ufu.br/handle/123456789/36092; http://doi.org/10.14393/ufu.di.2022.528

Breast cancer is the second most common type of cancer that affects woman in Brazil. Early detection of this disease is essential for improving chances of cure and recovery of the patients. Infrared thermography is an imaging technique being studied that has shown good potential in the early detection of breast cancer. In this work, we use three pre-trained convolutional neural networks (CNNs) VGG-16, ResNet-50 e DenseNet-201 to classify breast static thermography images (sick or healthy). Finding suitable CNN archi- tecture and hiper-parameters is essential for a good classification outcome. Furthermore, this is not an easy task. In order to address this problem of finding suitable architecture and hiper-parameters for the fully connected part of used CNNs, we used two bio inspired optimization techniques: genetic algorithm (GA) and particle swam optimization (PSO). Using the optimization techniques we were able to improve the F1-score from 0.66 to 0.92 for the VGG using the GA and the F1-score of the ResNet from 0.83 to 0.90 using the GA and PSO. Moreover, the optimization also maintained the good results of 0.92 F1-score for the DenseNet. Finally, we opt for using a random forest surrogate model in order to obtain a cheaper model. Using the surrogate, we improved the DenseNet from 0.92 to 1 of F1-score.