Automated cup-to-disc ratio evaluation for generalised glaucoma diagnosis using deeplabv3+

Glaucoma, a neurodegenerative disease is the leading cause of irreversible vision loss, where the intraocular pressure in the retina builds up due to the blockage of the drainage system (trabecular meshwork) between iris and cornea because of increased aqueous humor fluid. This damages the optic nerve head (ONH) which is responsible for carrying information from the eye to the brain, resulting in the thinning of the retinal nerve fibre layer and change of shape of ONH. Often, it is difficult for the patients to realize the loss of field of vision from the periphery regions towards the centre. Therefore, to restrict the impediments of the disease, it demands an early and accurate detection method, required for timely intervention and treatment. In this paper, we focus on producing a primary assistance cup-to-dise ratio (CDR) report for doctors in glaucoma assessment with the help of a deep learning segmentation model (Deeplabv3+ with ResNet-50). The model retains the spatial information of the contextual features which in turn results in increased segmentation accuracy. A database containing 3138 retinal eye fundus images which draws data from varied ethnicities around the globe is selected for the model training to increase the adaptability of the model during prediction. A validation intersection over union (IOU) of 94.6% and 85% was obtained for optic disc and cup segmentation respectively. The model achieved an IOU of 96.2% and 88.31% on the test data kept separately from the training data and an IOU of 90% and 74% on a new clinical anonymised patient data collected from the hospital. The results show how well the model has generalized when tested on a new dataset and the comparability of the model with the other state-of-the-art (SOTA) methods.