Deep Learning Framework for Early Glaucoma Diagnosis: ResNet Classification & YOLOv5 Segmentation

The blinding condition known as glaucoma proves to be the main contributor to permanent blindness across the world while simultaneously acting as a significant ocular health issue. Early detection of vision-threatening risks serves as the foundation to address the issue until vision loss is prevented from progressing. Fundus image diagnosis using human eye examination as performed by ophthalmologists remains a challenging method because it is time-consuming and subjective and error-prone. A new deep learning framework designed with two sequential phases allows the identification followed by segmentation of glaucoma elements found within fundus images. A ResNet50 model operates in the first phase to distinguish between glaucoma-positive and -negative images and it eliminates cases without glaucoma before human review. The current high-end object detection algorithm YOLOv5 runs in the second stage to precisely segment optic disc and cup structures in images that indicate glaucoma. Automatic measurement of the essential glaucoma diagnostic factor known as the cup-to-disc ratio (CDR) becomes possible through segmentation. Public database tests demonstrated that the proposed framework achieved classification accuracy at 94.2% while the segmentation results showed Dice coefficients at 0.91 for optic disc and 0.87 for optic cup. The system possesses the capability to enhance clinical operations at ophthalmic practices by enabling through examinations to diagnose diseases at an early stage thus benefiting patients more effectively.