Glaucoma Detection Using Intelligent Analysis of Fundus Images: AI Pipeline, Mobile Application, and Low-Cost Fundus Camera Prototype

Abstract

Glaucoma is the second leading cause of irreversible blindness worldwide, affecting over 80 million people, with projections reaching 111 million by 2040. Early detection is critical, yet up to 50% of patients remain undiagnosed. This paper presents an integrated system for automated glaucoma screening combining a deep learning image analysis pipeline, an iOS mobile application, and a low-cost fundus camera prototype. The AI pipeline follows a two-stage architecture: a YOLOv9 model for region-of-interest (ROI) detection, followed by a UNet++ segmentation model for optic disc and cup delineation. The Cup-to-Disc Ratio (CDR) is computed from the resulting masks and used to classify glaucoma risk. Five YOLO model variants (v8, v9, v11, v12, v26) were evaluated on an augmented dataset of over 6100 fundus images; YOLOv9 achieved the best overall balance with precision of 98%. UNet++ reached approximately 90% accuracy on the segmentation task. The iOS application, built in Swift with an MVVM architecture and a FastAPI backend hosted on Microsoft Azure, streams real-time analysis progress to prevent frozen-screen effects. The fundus camera prototype, constructed from a Volk 20D lens and a PVC tube mounted on an iPhone 14 Pro, enables retinal image capture at a cost below 1300 PLN. The system targets clinical screening workflows while remaining accessible to individual users.