Advanced glaucoma detection through retinal fundus image segmentation and stacking classifier

Abstract Identification of diseases affecting the optic nerve is paramount in ophthalmology. Traditional diagnostic methods often rely heavily on manual intervention, leading to potential errors and delays. Consequently, there’s a pressing need for automated systems capable of early disease detection to initiate timely treatment, thereby averting permanent vision impairment. Glaucoma, a chronic eye condition with the potential for irreversible blindness, stands as a prime example. Recent strides in deep learning have revolutionized the diagnosis of glaucoma through the analysis of retinal fundus images. These technologies enable the development of algorithms capable of identifying subtle signs indicative of glaucoma, often imperceptible to the human eye. By harnessing vast datasets, these algorithms can learn to distinguish between healthy and diseased retinas with remarkable accuracy, facilitating early intervention and treatment. The novelty of the work comes from the combination of existing machine learning techniques on tabular datasets as well as deep learning image segmentation to increase the performance compared to using any one technique alone. The proposed methodology for glaucoma diagnosis makes use of a combination of a deep learning model and several machine learning and achieves Receiver Operating Characteristics Area Under Curve (ROC-AUC) of 88% (0.88) on healthy, glaucomatous and suspicious images in the testing data of the publicly available PAPILA dataset. The mean absolute error for the horizontal cup to disc ratio of 0.093 was obtained and the mean absolute error for the vertical cup to disc ratio of 0.082 was obtained on the testing dataset.