Métodos Computacionais para Análise e Classificação de Displasias em Imagens da Cavidade Bucal

• Ano de defesa: 2019
• Autor(a) principal: Silva, Adriano Barbosa
• 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: Computação; Displasia; Imagens Histológicas; Segmentação Nuclear; Redes Neurais Convolucionais; Classificador Polinomial; Processamento Digital de Imagens; Dysplasia; Histological Images; Nuclei Segmentation; Convolutional Neural Networks; Polynomial Classifier; Digital Image Processing; CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO
• Link de acesso: https://repositorio.ufu.br/handle/123456789/27114; http://dx.doi.org/10.14393/ufu.di.2019.2390

In recent years, the use of computer systems as a tool for diagnostic assistance has shown significant growth in applications aimed at histological tissues analysis. Computational algorithms are used to extract information that allows the quantification of lesions. This simplifies the manual diagnostic process performed by a specialist, which requires a lot of time, energy and is subject to subjectivity factors. In order to improve the diagnostic process of oral dysplasias, this work proposes a method for segmentation and classification of nuclear structures present in images of histological tissues. The proposed method is divided into the stages of segmentation, post-processing, feature extraction and classification. In the segmentation stage, the neural network Mask R-CNN was used to identify significant information to separate cell nuclei from background region. In the post-processing stage, dilation, region filling, and erosion operations were used to fill incomplete nuclei regions and remove remaining noise from the segmentation stage. In the feature extraction stage, texture and morphologic attributes were extracted from the images nuclei. Finally, a polynomial classifier algorithm was used to classify the images among healthy tissue, mild dysplasia, moderate dysplasia and severe dysplasia. The obtained results were compared with the groundtruth generated by a specialist and with other methods present in literature. The method obtained accuracy of 89.52% in the segmentation of the nuclear components, which is 14% higher than other methods. During the classification stage, the classification method obtained a value of area under the ROC curve of 0.92, a value 6.5% higher than other methods. The method obtained results more relevant than other methods in literature, showing that it can be used by health specialists as a tool to study pre-cancer lesions.