Deep Learning & Digital Fundus Images: Glaucoma Detection using DenseNet

Being one of the most serious causes of irreversible blindness, all over the world, glaucoma is the asymptomatic disease which produces damage at the optic nerve head level. The symptoms related to glaucoma are not noticeable until severe stages of this illness and they appear when the patient has already lost a significant part of his eyesight. As the detection of glaucoma is vital for preventing blindness, regular eye investigations have to be made. However, these examinations might not be relevant for glaucoma diagnosis. Nonetheless, with the continuous development of deep learning algorithms, early detection of glaucoma based on digital fundus images, is possible. As convolutional neural networks (CNNs) have proven good performance in early detection of several diseases, they have been recently used in the ophthalmological field for the identification of certain eye illnesses, including glaucoma. This paper proposes a new method which uses densely connected neural networks (DenseNet) with 201 layers, initially pre-trained on ImageNet, using ACRIMA dataset. An accuracy of approximately 97% and an f1-score of 0.969 were obtained which gives us cause for careful optimism concerning the usefulness of this classification model for the early detection of glaucoma.