Aprendizagem Profunda Aplicada ao Diagnóstico de Melanoma

• Ano de defesa: 2019
• Autor(a) principal: MAIA, Lucas Bezerra
• Orientador(a): BRAZ JÚNIOR, Geraldo
• Banca de defesa: BRAZ JÚNIOR, Geraldo , PAIVA, Anselmo Cardoso de , ALMEIDA, João Dallyson Sousa de , CARVALHO FILHO, Antonio Oseas de
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal do Maranhão
• Programa de Pós-Graduação: PROGRAMA DE PÓS-GRADUAÇÃO EM CIÊNCIA DA COMPUTAÇÃO/CCET
• Departamento: DEPARTAMENTO DE INFORMÁTICA/CCET
• País: Brasil
• Palavras-chave em Português: Diagnóstico de melanoma; Balanceamento de classes; Aprendizagem profunda; Redes neurais convolucionais; Random forest; SMOTE; ENN
• Palavras-chave em Inglês: Melanoma diagnosis; Imbalanced learning; Deep learning; Convolutional neural networks; Random forest; SMOTE; ENN
• Área do conhecimento CNPq: Teoria da Computação; Análise de Algoritmos e Complexidade de Computação; Cancerologia
• Link de acesso: https://tedebc.ufma.br/jspui/handle/tede/2578

Melanoma is the most lethal type of cancer when compared to others skin diseases. However, when the diagnosis is made in its initial stage, patients have high rates of recovery. Several approaches to automatic detection and diagnosis of melanoma have been explored by different authors in order to provide an auxiliary opinion to specialists. Training models with the existing data sets have been a difficult task due to the problem of imbalanced data. This work aims to evaluate to the evaluation the performance of machine learning algorithms combined with imbalanced learning technique, regarding the task of melanoma diagnosis. The architectures of Convolutional Neural Networks VGG16, VGG19, Inception, and ResNet were used along with ABCD rule to extract patterns of skin lesions in a set of 200 dermatoscopic images. The Random Forest classifier reached a sensitivity of 92.5 % and a kappa index of 77.15 % after the use of attribute selection with Greedy Stepwise and balancing the training data with Synthetically Minority Oversampling TEchnique (SMOTE) and the Edited Nearest Neighbor (ENN) rule.