Classificação de displasia da cavidade oral baseada em descritores fractais e ensemble learning

• Ano de defesa: 2024
• Autor(a) principal: Carvalho, Rafael Henrique de Oliveira
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso embargado
• Idioma: por
• Instituição de defesa: Universidade Federal de Uberlândia; Brasil; Programa de Pós-graduação em Ciência da Computação
• Palavras-chave em Português: Comitês de classificadores; Displasia; Imagens Histológicas; Redes Neurais Convolucionais; Processamento Digital de Imagens; Ensemble Learning; Dysplasia; Histological Images.; Digital Image Processing; Convolutional Neural Networks; CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO::METODOLOGIA E TECNICAS DA COMPUTACAO; Computação; Redes neurais (Computação); Fractais; Boca - Câncer; ODS::ODS 3. Saúde e bem-estar - Assegurar uma vida saudável e promover o bem-estar para todos, em todas as idades.
• Link de acesso: https://repositorio.ufu.br/handle/123456789/43129; http://dx.doi.org/10.14393/ufu.di.2024.5077

Lesions in the oral cavity can be classified into different grades by specialists. However, due to workload and level of experience, this task may be subject to subjectivity. One way to assist the specialist in this task in recent years has been the use of computational systems as supplementary tools for decision-making. In this work, an approach is presented to classify lesions in the oral cavity based on descriptors obtained from fractal geometry, models of convolutional neural networks, and ensemble learning. In the first step, the gliding-box algorithm was applied to obtain local descriptors of fractal geometry in a multiresolution analysis. This attribute vector was reshaped into a 2D matrix using representations based on recurrence plot, Markov transition field, and self-similarity matrix. The 2D matrices were used as input for the ResNet-50 and EfficientNet convolutional networks. In this study, the impact of optimizing the fully connected layers of these structures was investigated. Finally, ensemble learning models of classifiers with the sum rule were applied to the datasets. The efficiency of the proposed methodology was analyzed on a dataset composed of 74 regions of interest in oral epithelial dysplasia categorized into healthy, mild, moderate, and severe classes. The proposed method used cross-validation technique and achieved accuracy rates of up to 98.6%. This approach becomes a complementary tool for application in clinical practice, aiding the specialist in decision-making regarding lesion classification.