AI-Based Eye Disease Detection from Fundus Images

Problem statement

Eye diseases like diabetic retinopathy and glaucoma can lead to permanent vision loss if not detected early. Many patients, especially in rural areas, do not have access to frequent eye screenings. Automating the initial screening of fundus images can help identify at-risk patients and reduce the load on ophthalmologists.

Abstract

This project aims to build a deep learning–based screening system that can classify retinal fundus images as healthy or diseased. Using publicly available datasets of labeled eye images, a convolutional neural network is trained to detect signs of diabetic retinopathy or other abnormalities such as hemorrhages and exudates. A preprocessing pipeline enhances contrast and normalizes images. A simple web interface allows uploading fundus images and returns a risk level or disease class, making it a potential tool for tele-ophthalmology.

Components required

• Dataset of retinal fundus images (e.g., Kaggle DR, Messidor)
• Python with TensorFlow/Keras or PyTorch
• OpenCV for image preprocessing
• GPU-enabled PC or cloud instance (optional but recommended)
• Web UI (Flask / Django / React frontend)

Block diagram

Working

The user or technician uploads a retinal image captured by a fundus camera. The system resizes, normalizes and enhances the image to improve contrast. The processed image is passed through a trained CNN which outputs probabilities for different disease classes or severity levels. Based on thresholds, the system flags images that need urgent review and displays a simple report to the user, which can be shared with ophthalmologists for confirmation.

Applications

• Telemedicine and remote eye screening camps
• Primary health centers with limited specialist availability
• Hospital pre-screening tools
• Research in medical image analysis