Early Detection of Glaucoma Using Residual Networks

Glaucoma is one of the leading causes of irreversible blindness around the world and remains asymptomatic until its later stages. Therefore, early diagnosis is of crucial importance. The detection of glaucoma in early stages (from color fundus images) is a challenging task, since the clinical signs in the retinal images are very subtle and go undetected most of the time by the human eye. Convolutional neural networks have proven to provide good results for automatic detection of subtle features from images. In this work we explore the possibility of using residual networks to detect early stages of glaucoma. We introduce a proprietary early-stage glaucoma fundus color images dataset. We used a ResNet50 network which initially was trained on the ImageNet dataset. The level of accuracy on the validation set was 96.95%. The results indicate that using deep learning algorithms a cost-effective screening tool could be built for early and costeffective detection of glaucoma.