Glaucoma Detection through Deep Learning: A Transfer Learning Techniques using CDR Feature Extraction

Elevated intraocular pressure is the hallmark of glaucoma, which can cause vision loss and optic nerve damage if initially undiagnosed due to mild symptoms. For vision impairment to be effectively managed and prevented, early detection is essential. Optical coherence tomography and fundus imaging are used in traditional diagnosis; thus, the optic cup-to-disk ratio is a crucial parameter that needs precise segmentation. This work uses the ensemble technique of ResNet101, MobileNet, and a refined VGGNet-19 model for subtle structural learning from RGB fundus images and spatial coordinates. The approach addresses challenges related to glaucoma categorization as well as optic cup/disk segmentation. Overall accuracy on the test set was 94%. VGG19 scored 0.92, 0.95, and 0.93 F1-score, while ResNet101 scored 0.93, 0.94, and 0.935, and MobileNet scored a competitive 0.915 F1-score. This work highlights the clinical potential and significance of reliable early screening for the management of this vision-threatening disorder by demonstrating how segmentation and deep learning approaches can be used to enable effective automated glaucoma diagnosis.