Exhaustive Feature Selection Using Wrapper method for Artery-Vein Classification in Retinal Fundus Image

Retinal imaging plays a crucial role in diagnosing ocular and systemic disorders by analyzing abnormalities in retinal blood vessels. Numerous diagnostic parameters, such as vascular caliber, centerline pixels, vessel bifurcation, arteriolar constriction, and color reflections, have been explored using machine learning and deep learning approaches for artery-vein classification. While these methods have demonstrated high-performance outcomes, limited research has investigated alternative approaches to identify optimal features for artery-vein classification. This study addresses this gap by evaluating nine color features using a wrapper method across three benchmark datasets: DRIVE, HRF, and VICA VR. The DRIVE dataset achieved the best classification performance, with an Area Under the Curve of 91.49%, accuracy of 83.10%, sensitivity of 84.68%, and specificity of 81.47%. Among the features studied, the Chrominance-Blue (Cb) and Chrominance-Red (Cr) components emerged as the most influential. These findings underscore the critical role of color features in enhancing artery-vein classification. They provide a foundation for further research into feature optimization, potentially improving diagnostic tools for retinal and systematic diseases.