A Capsule Network-based for identification of Glaucoma in retinal images

Glaucoma is an eye disease responsible for the second most common cause of blindness in the world. The need to detect this disease in its early stages is notorious, considering that late treatment can cause loss of vision. In this context, computational methods are being developed to assist specialists in the task of analyzing ocular images to provide greater precision to the diagnosis. In this paper, we present a methodology for automatic classification of glaucoma using Capsule Network (CapsNet), a recent model of deep learning that analyzes the hierarchical spatial relationships between characteristics to represent images, so that it requires fewer training samples than traditional CNNs to achieve efficient classification. Before the execution of CapsNet, we applied a pre-processing step to the images, in order to highlight the characteristics. Our results were promising, with 90.90% accuracy, 86.88% recall, 94.64% precision, 90.59% f1-score, 0.904 AUC and 0.801 kappa index. The main contribution of our method is the fact that we have achieved promising results without the need to apply data to increase and segment the region of the optical disc. Thus, our study showed the potential of the capsules in identifying the relationships between the characteristics, even in the face of a reduced set of training.