Abstract

Applied Artificial Intelligence: Differentiating Between Glaucomatous, High- and Low-Risk Suspect, and Normal Eyes with Neural Networks

Though glaucoma is the leading cause of worldwide irreversible blindness, few studies have been performed on using artificial intelligence (AI) to diagnose it. It was hypothesized that it is possible for convolutional neural networks (CNNs) to be trained to surpass the current best accuracy due to AI’s increasing influence in ophthalmology and its self-correcting nature. 2811 color fundus images were collected from a variety of sources: Drishti-GS, RIM-ONE, Harvard Dataverse, and the Rand Eye Institute. A CNN was trained for 100 epochs, with a random 80% of images used for training and the remaining images for evaluation. The accuracy of the CNN after training was measured on previously unseen images. Additionally, to increase replicability of the results, a 10-times 10-fold cross validation was performed. The accuracy of the CNN trained for 100 epochs was 98.93%. The average accuracy after the cross validation was 92.53%. Though AI has shown success in binary classification problems, accuracy significantly decreases when more classes are added. Given that this research trained a CNN to differentiate between 4 categories with no statistically significant difference with perfect performance and a statistically significant difference by chi square tests with the current best glaucoma-diagnosing CNNs (88.05%), there is enormous potential for AI to increase awareness for glaucoma and access to medical care, helping doctors improve lives.