Aiding Glaucoma Diagnosis from the Automated Classification and Segmentation of Fundus Images

Glaucoma is the most significant cause of irreversible vision loss and the second biggest cause of blindness globally. The first signs of the disease will only appear in an advanced stage when there is no more recovery. Early diagnosis is of the utmost importance and is currently performed primarily through fundoscopy. This fundus image exam is a tedious and manual process, prone to human errors that can result in false negatives, which could promote vision loss. Deep learning approaches are being used to detect glaucoma directly from eyes fundus images, achieving promising results. However, there is no apparent interest in deploying these systems in medical clinics, which would require lightweight models with minimal false negatives and comprehensive outputs. The present study explores these gaps and proposes an architecture composed of one segmentation network for disc and cup and one classification network for direct glaucoma classification. Our main contribution is optimizing a glaucoma-aiding system by increasing model sensitivity by 3% and simplifying the architecture. Unlike related works, we present a lighter model that shows valuable information to the physician, building their trust in the system.