Auxílio ao diagnóstico automático do esôfago de Barrett utilizando aprendizado de máquina

• Ano de defesa: 2022
• Autor(a) principal: Souza Júnior, Luis Antonio de
• Orientador(a): Papa, João Paulo
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Universidade Federal de São Carlos; Câmpus São Carlos
• Programa de Pós-Graduação: Programa de Pós-Graduação em Ciência da Computação - PPGCC
• Palavras-chave em Português: Aprendizado de máquina; Esôfago de Barrett; Aprendizado em profundidade; Redes neurais convolucionais; Interpretabilidade
• Palavras-chave em Inglês: Machine learning; Barrett’s esophagus; Deep learning; Convolutional neural networks; Interpretability; Handcrafted features; Deeply-learnable features
• Área do conhecimento CNPq: CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO::METODOLOGIA E TECNICAS DA COMPUTACAO::LINGUAGENS DE PROGRAMACAO
• Link de acesso: https://repositorio.ufscar.br/handle/20.500.14289/15820

Esophageal adenocarcinoma is an illness that is usually hard to detect at the early stages in the presence of Barrett's esohagus. The development of automatic evaluation systems of such illness may be very useful, thus assisting the experts in the neoplastic region detection. With the strong growth of machine learning techniques aiming to improve the effectivess of medical diagnosis, the use of such approaches characterizes a strong scenario to be explored for the early diagnosis of esophageal adenocarcinoma. Barrett's esophagus as a predecessor of adenocarcinoma can be explained by some risk factors, such as obesity, smoking, and late medical diagnosis. This project proposes the development of new computer vision and machine learning techniques to assist the automatic diagnosis of the esophageal adenocarcinioma based on the evaluation of two kind of features: (i) handcrafted features, calculated by means of human knowledge using some image processing technique and; (ii) deeply-learnable features, calculated exclusively based on deep learning techniques. From the extensive application of global and local protocols for the models proposed in this work, the description of cancer-affected images and Barrett's esophagus-affected samples were generalized and deeply evaluated using, for example, classifiers such as Support Vector Machines, ResNet-50 and the combination of descriptions by handcrafted and deeply-learnable features. Also, the behavior of the automatic definition of key-points within the evaluated techniques was observed, something of a paramount importance nowadays to guarantee transparency and reliability in the decisions made by computational techniques. Thus, this project contributes to both the computational and medical fields, introducing new classifiers, approaches and interpretation of the class generalization process, in addition to proposing fast and precise manners to define cancer, delivering important and novel results concerning the accurate identification of cancer in samples affected by Barrett's esophagus, showing values ​​around 95% of correct identification rates and arranged in a collection of scientific works developed by the author during the research period and submitted/published to date.