Early Glaucoma Detection using Deep Learning with Multiple Datasets of Fundus Images

Glaucoma is an eye condition that damages the optic nerve, which can lead to vision loss or blindness. This condition affects individuals worldwide, but early glaucoma detection can help diagnose the condition faster and enhance patient treatment. Traditional diagnostic methods, such as Tonometry, Ophthalmoscopy, and Gonioscopy are costly, invasive to the eye, and require a medical specialist. However, non-invasive methods such as deep-learning approaches based on fundus images of the eye show promising results but such architectures are typically trained on single datasets, which limits their practical generalizability to different patients. In this project, we develop a convolutional neural network (CNN) model based on the EfficientNet architecture, trained sequentially across the ACRIMA, ORIGA, and RIM-ONE datasets of fundus images, to enhance diagnostic accuracy and model generalizability. By conducting experiments on the trained model and evaluating metrics such as accuracy, sensitivity, specificity, and AUC-ROC, we demonstrate this method's capability for improved glaucoma detection and its potential use in clinical data for early detection. Ultimately, our work aims to deliver an accurate, easy-to-use, and scalable model for non-invasive early glaucoma screening, which contributes to better patient treatment through timely clinical intervention.