Automated Diabetic Retinopathy Detection via U-Net Segmentation and CNN Classification

Abstract

Diabetic Retinopathy (DR) is a sight-threatening condition that requires early diagnosis to prevent vision loss. This project proposes a novel deep learning-based system combining U-Net segmentation and Convolutional Neural Networks (CNN) for efficient DR detection from retinal fundus images. The U-Net model segments key retinal regions, preserving crucial pathological features through a region-merging technique. These refined images are classified into five DR severity stages using CNN. The proposed hybrid approach addresses challenges like image variability and lesion misdetection, achieving an accuracy of 93.33% on public datasets including DRIVE and Kaggle DR. This system offers scalable, automated screening with high precision, aiding ophthalmologists in early diagnosis and improving healthcare accessibility in resource-limited settings.