An Enhanced Segmentation and Deep Learning Architecture for Early Diabetic Retinopathy Detection
Gattu Tejaswini,
K Dhivya Shrie,
Akhil G,
K Varshitha
Diabetic retinopathy (DR) is a leading cause of preventable blindness among diabetic patients,
necessitating early detection through regular screening of retinal fundus images. Traditional manual
diagnosis by ophthalmologists is time-consuming, subjective, and resource-intensive, often leading to
delayed intervention. This paper proposes an enhanced segmentation and deep learning architecture
that integrates advanced image preprocessing, lesion segmentation using an optimized U-Net variant,
and multi-stage classification via convolutional neural networks (CNNs) with transfer learning. The
system automatically detects and segments key DR lesions (microaneurysms, hemorrhages, hard/soft
exudates) at the pixel level, followed by severity grading into five classes (No DR, Mild, Moderate,
Severe, Proliferative). A hybrid approach combines spatial feature extraction with attention mechanisms
to improve lesion localization in early stages. Experimental evaluation on benchmark datasets (e.g.,
APTOS, EyePACS, DDR) demonstrates superior segmentation accuracies of 87.10% IoU and 84.50%
Dice coefficient, alongside classification accuracy of 99.20% in binary referable DR detection and
96–98% in multi-class grading. The proposed framework offers cost-effective, scalable, and portable
screening, reducing diagnostic delays and enhancing accessibility in resource-limited settings while
maintaining high sensitivity for early-stage identification.
