Glaucoma Disease Detection System Using Hybrid Deep Learning Architecture

Abstract

Glaucoma is one of the major causes of permanent loss of vision in the world. Blindness can best be prevented by mass screening of the disease at its early stages. Nonetheless, the manual checking of eye photos is time consuming, and it is also vulnerable to human error. This paper describes a hybrid deep learning architecture that can be used to detect glaucoma by examining eye photographs. The suggested system isolates the procedures of locating the eye anatomy as well as categorizing the disease to simplify the outcomes and make them easy to comprehend. First, an Attention U-Net masks Optic Disc and Optic Cup to produce accurate black and white masks. The masks are added to the original color image to form a 5-channel input. The resulting combined input is then used by a Data-efficient Image Transformer (DeiT-Tiny) that then determines the physical distances between the eye structures. The hybrid architecture was tested on the SMDG-19 dataset of 12,449 images with accuracy of 89.77%, sensitivity of 84.41% and Area Under the Curve (AUC) of 0.9564. The results of the experiment indicate that the local feature extraction and global classification should be separated to produce an extremely useful and effective medical screening program tool. This study specifically focuses on glaucoma detection rather than general retinal diseases.