Detecção de cárie dentária por fotoluminescência utilizando processamento digital de imagem com inteligência artificial

• Ano de defesa: 2021
• Autor(a) principal: Carneiro, Davi Clementino
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal da Paraíba; Brasil; Odontologia; Programa de Pós-Graduação em Odontologia; UFPB
• Palavras-chave em Português: Diagnóstico bucal; Inteligência artificial; Processamento de imagem assistida por computador; Oral diagnosis; Artificial intelligence; Computer-assisted image processing; CNPQ::CIENCIAS DA SAUDE::ODONTOLOGIA
• Link de acesso: https://repositorio.ufpb.br/jspui/handle/123456789/22777

Dentistry is among the health sciences that is most dependent on technology. The advancement of processing capabilities has enabled the insertion of several computational methods in different medical areas. Among the instruments to assist in decision making, digital diagnosis has been shown to be a good alternative, generating more reliable outcomes, increasing practicability and being quicker as results to the incorporation of technology. The objective was to develop an automated system for detecting initial caries lesions from photographs with fluorescent light using artificial intelligence. For this study, 67 oral photographs from different angles and approaches were taken of 18 patients between 8 and 14 years old. These images were filtered and cut into 691 final samples that comprised this research. The samples were segmented using the RGB, HSB/HUE and Grayscale color systems, without modifying them. An Support Vector Machine (SVM) algorithm and convolutional neural networks were used to train the system in two generations under Python and C++ coding. The results presented by the threshold_li algorithm in association with the green channel proposed better dissociation of structures in the same image. The expansion of the sample showed a significant improvement in software analysis after the insertion of the second generation of data. The automated detection of initial demineralization presented a satisfactory result in comparison with the gold standard established in singular elements and in photographs of full mouth, as related in a learning curve, in addition to graphs of dissimilarity and scalability. Even greater insertion of data and improvements in the software code can make the system even more intelligent, providing more accurate and more reliable responses.