Modelo de aprendizado de máquina com fronteira de decisão fechada para identificação de carcinoma hepatocelular inicial

• Ano de defesa: 2018
• Autor(a) principal: Thomaz, Ricardo de Lima
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Uberlândia; Brasil; Programa de Pós-graduação em Engenharia Elétrica
• Palavras-chave em Português: Carcinoma hepatocelular; Hepatocellular carcinoma; Detecção automática; Automatic detection; Classificação; Classification; Classificadores de uma classe; One-class classifiers; Sistemas CAD; CAD systems; Engenharia elétrica; Electrical engineering; Fígado; Liver; Câncer; Diagnóstico; Tomografia computadorizada; Computed tomography; CNPQ::ENGENHARIAS::ENGENHARIA BIOMEDICA::BIOENGENHARIA::PROCESSAMENTO DE SINAIS BIOLOGICOS
• Link de acesso: https://repositorio.ufu.br/handle/123456789/21274; http://dx.doi.org/10.14393/ufu.te.2018.769

Hepatocellular carcinoma is the fifth most common and the third most deadly type among all cancers. Its early detection is of high relevance since curative treatments can be used. Thus, computer systems available in the literature aim to aid in the diagnosis of hepatic lesions in computed tomography. Among the methods for automatic detection, the use of autoencoders for characterizing the hepatic parenchyma and searching lesions show good results. However, this search will not necessarily result in the detection of lesions. Thus, we hypothesized that the use of a one-class model that characterizes early HCC as a fitness function of a metaheuristic will allow the identification of lesions in CT scans. For this, we proposed the use of an ensemble of one-class support vector machines (EOCSVM) to characterize early HCC. By using these models as the fitness function of a Particle Swarm Optimization algorithm, regions with a high probability of early HCC were detected. After evaluating 28 exams by a cross-validation method, 97.05% of the lesions were correctly detected by the EOCSVM model that prioritized the minimization of false negatives. Models prioritizing F-Score and false positives resulted in 89.70% and 82.35% of lesions detected, respectively. By evaluating the results at the point of maximum sensitivity, the EOCSVMFN resulted in an average of 3.91 false positive cases per exam, similarly to the state-of-the-art. The areas under the ROC curves for the models prioritizing FN, F-Score and FP resulted in 0.94, 0.84 and 0.82, respectively. Therefore, the results allowed us to confirm the hypothesis that the use of an OCSVM model characterizing early HCC as a PSO fitness function allows the detection of early HCC in CT exams. As the main scientific contribution, this thesis introduced the use of closed decision boundary models as fitness function of an optimization algorithm with the focus on detecting regions of interest.