Lightweight Attention-Enhanced U-Net Framework for Automated Multi-Class Diabetic Retinopathy Lesion Segmentation Using Retinal Fundus Image

Abstract

Diabetic Retinopathy (DR) is a major diabetes-related eye disease that can cause permanent blindness if not diagnosed early. This paper presents a lightweight CNN-based U-Net framework for automated multi-class segmentation of retinal lesions from fundus images. The proposed system uses image preprocessing, data augmentation, and hybrid loss functions to improve segmentation accuracy and handle class imbalance. The model effectively detects lesions such as microaneurysms, hemorrhages, hard exudates, soft exudates, and optic disc regions using the IDRiD dataset. Experimental results demonstrate reliable performance in terms of Dice score, IoU, precision, and recall, supporting early diagnosis and clinical decision-making.