Validação de redes generativas condicionadas para segmentação de escavação óptica em imagens de fundo de retina

Abstract

The evaluation of retinal fundus images represents an important task in ophtalmology, as it provides indication of eye-related pathologies. Among them, glaucoma stands out due to the need of an early diagnosis and early treatment, so severe vision symptoms can be avoided. Due to the high cost and low availability of retina specialists, automatic processes that identify adverse structures in retinal fundus images can aid the process of obtaining diagnoses. For glaucoma diagnosis, it is important to identify the dimension of the optic cup in the retina nerve head, as a ratio between optic cup and optic disc greater than 0.5 is a strong indicator of the condition. An automatic process depends on the segmentation between optic disc and optic cup in retinal fundus images, which provides the structures dimensions for calculating such ratio. This work proposes the usage of conditional generative adversarial neural networks for the retina segmentation task, based on Pix2Pix architecture. To validate the model, the proposed model was compared to U-Net and M-Net convolutional models, that represent the best results in literature. Results indicate that the generative model is capable of providing retina segmentation with precision comparable with state-of-art models, and it is capable of doing such task with higher robustness.