Análise e comparação do desempenho de diferentes arquiteturas de redes neurais artificiais profundas aplicadas á triagem de lesões de pele

• Autor(a) principal: Santos, André Pires dos
• Data de Publicação: 2019
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: Biblioteca Digital de Teses e Dissertações do Mackenzie
• Texto Completo: http://dspace.mackenzie.br/handle/10899/24505
• topic: redes neurais profundas; visão computacional; lesões de pele; dermatologia; CNPQ::ENGENHARIAS::ENGENHARIA ELETRICA

This work aims to study the performance of several deep neural network architectures applied to classify dermatological lesions. The result has as secondary applications to set the referral and priority of care to the patient, increasing the efficiency to deliver the appropriate treatment for patients lacking access to this medical specialty. The study presents a brief introduction of the main reasons, the historical origin and the impact from the lack of access of the population to the health system. The work also provides the potential benefits from the insertion of artificial intelligence in the health care processes. It also presents a review of scientific studies aimed at the classification of skin lesions by artificial intelligence algorithms and the main concepts and history of artificial neural networks from the beginning to the detailed description of the most recent architectures. As results, the work presents the results of two experiments that together contemplate different scenarios of the application of deep neural networks to measure the accuracy of the proposed algorithms applied to classify skin lesions. The experiments evaluate the performance several architectures, varying parameters and weight initialization strategies. As a main result, the work presents that the GoogLeNet architecture trained with 24,000 images, during a thousand epochs using random initialization as strategy to initialize weights and learning rate of - , was able to obtain an accuracy of 89.72% for diagnostic suggestion, 92 % for priority indication and 96.03% for patient referral in a 6,975 image test set.