Diabetic retinopathy is a major cause of blindness worldwide, and accurate retinal vessel extraction is essential for its early detection and monitoring. Manual delineation by ophthalmologists is time-consuming and susceptible to human variability, while automated segmentation remains challenged by noise, illumination changes, and weak contrast in thin vessels. This study presents a novel hybrid multi-stage approach for automated blood vessel extraction from digital retinal images. The method integrates noise reduction, multi-directional derivative–based centerline detection, and region-expansion reconstruction with noise suppression to more reliably recover fine and low-contrast vascular structures commonly missed by existing techniques. The algorithm was evaluated on the DRIVE dataset, achieving an accuracy of 0.96578, sensitivity of 0.92896, and specificity of 0.98965. Compared with leading classical and deep-learning baselines, the proposed framework achieves approximately 0.77% higher accuracy and over 1% improvement in thin-vessel sensitivity. These quantitative gains confirm the effectiveness of the hybrid stages in enhancing vessel continuity and recall. The proposed method therefore offers a robust and efficient tool for automated retinal image analysis, with potential utility in supporting early diagnosis and clinical management of diabetic retinopathy in practical screening workflows.
Afrakhteh et al. (2026) studied this question.