The retina, a transparent neural tissue lining the eye, is crucial for detecting both ocular and systemic diseases through imaging techniques such as fundus photography, OCT, and angiography. Recent advancements in artificial intelligence and machine learning have significantly improved retinal image analysis, enabling automated classification of arteries and veins to assists in diagnosing conditions like diabetic retinopathy, glaucoma, and systemic diseases such as hypertension and Chronic Kidney Disease (CKD). However, pixel-level analysis of retinal features, particularly colour features, remains underexplored, with challenges such as image variability and non-linear feature relationships hindering optimal classification. This study addresses these gaps by analyzing nine colour features—Red, Green, Blue, Hue, Saturation, Value, Y, Cb, and Cr—extracted pixel-wise from regions of interest (ROI) across three datasets: DRIVE, HRF, and VICAVR. A total of 3864, 9804, and 7466 normalized artery and vein pixel values underwent F-tests and two-sample T-tests, revealing statistically significant differences (p <0.05) for all features. The Minimum Redundancy Maximum Relevance (mRMR) algorithm ranked features by relevance, and combinations of top-ranked features were evaluated using nine machine learning classifiers. The Ensemble model, utilizing Random Forest Bagging with Decision Tree learners and optimized hyperparameters, achieved the highest AUC scores: 91% (DRIVE), 88.2% (HRF), and 90.1% (VICAVR). Incorporating all nine features yielded the best classification results, emphasizing their complementary roles. These findings demonstrate the potential of colour features in improving retinal vascular analysis, offering insights for non-invasive diagnostics and disease monitoring.
Saad et al. (Wed,) studied this question.