Estudo da aplicação de deep learning como auxílio ao diagnóstico de glaucoma

Abstract

Approximately 80 million people have glaucoma worldwide. It is estimated that half of them are unaware of their condition. In low and/or middle-income countries, more than 90% of people with glaucoma are still undiagnosed. Among those diagnosed, 35% are already blind. Affordable and effective screening approaches are therefore needed to identify the individuals at risk for vision loss. Widespread use of screening using fundus imaging, with AI-assisted classification, could allow glaucoma to be diagnosed alongside the other leading causes of blindness at low cost. Implementation studies are needed to determine how and where to apply these new tools. Many important research questions remain unresolved and require substantial investment and a concerted global effort to answer. Meeting the current need, the present research proposes to evaluate the application of deep learning methods in the diagnosis of glaucoma. To accomplish it, four existing CNN architectures (InceptionV3, SqueezeNet, VGG16 and VGG19) were applied to approximately 1550 fundus exams, evaluating different image treatments. In the best performance scenario, an accuracy of 0.71014, precision of 0.71019, recall of 0.71014 and F1 score of 0.71012 were observed. As next steps, it is planned to link the training results to other medical data, as well as other imaging tests in order to improve the method.