Segmentation and Detection of Diabetic Retinopathy Lesions: An Overview

Diabetic Retinopathy (DR) is a major cause of preventable blindness, making early detection of retinal lesions in fundus images critical. Our study provides an overview of the advances in DR lesion and anatomical-structure segmentation and detection, spanning from early image processing and traditional machine learning to recent deep learning (DL) methods. These methods are grouped by their learning paradigm and are then organized by targeted lesion type—red lesions (microaneurysms, hemorrhages), bright lesions (exudates), and anatomical structures (optic disc, vessels, fovea). In addition, we discuss datasets and evaluation metrics supporting this progress. While supervised DL architectures, especially U-Net variants with attention and transformers lead current performance, they remain constrained by annotation costs, domain shifts, and sensitivity to small lesions. Emerging unsupervised and partially supervised strategies, along with lightweight, multimodal, and uncertainty-aware approaches, aim to build more reliable and clinically relevant methods for early DR management.