Aprendizagem Profunda Aplicada ao Diagnóstico do Glaucoma

Abstract

Glaucoma is a cluster of ocular diseases that cause damage to the eye’s optic nerve and cause successive narrowing of the visual field in affected patients, due to an increase in intraocular pressure, which can lead the patient to blindness at an advanced stage without clinical reversal. For several years, from techniques of manual analysis of the internal structures of the eye to the use of deep learning with convolutional neural networks (CNNs) were successfully used in the diagnosis of glaucoma. However, building a deep learning network requires a lot of effort that in many situations is not always able to achieve satisfactory results due to the amount of parameters that need to be configured to adapt the CNN architecture to the problem in question. The objective of this work is to use a hyperparameter search technic to select the tuned parameters of a genetic algorithm (GA) to select the best CNN architecture through evolutionary techniques and to be able to aid in the accurate diagnosis of glaucoma, in eye fund images. The proposed methodology was applied in 455 images from RIM-ONE dataset, in its version 2 (r2), with resized images to 96x96 pixels in the RGB color model. The selected CNN by AG, after its training, achieved for the diagnosis of glaucoma the results of 96.63% for accuracy, 94.87% for sensitivity, 98.00% for specificity, 97.37% for precision and 96.10% for f-score.