Hybrid Feature Engineering and Sampling Strategies for Enhanced Diabetic Retinopathy Classification

This study presents a machine learning approach for Diabetic Retinopathy (DR) classification, integrating advanced preprocessing, feature extraction, and adaptive sampling. Preprocessing techniques, including CLAHE, gamma correction, and noise reduction, standardize image quality while preserving anatomical details. Class imbalance is addressed with ADASYN, generating synthetic samples in sparse regions for balanced class representation. Feature extraction focuses on clinical markers like vascular density, microaneurysms, exudates, and texture descriptors such as GLCM and LBP. Recursive Feature Elimination (RFE) refines features for better interpretability and reduced computational load. This binary classification framework achieves 92.91% accuracy on test data and a cross-validated ROC AUC of 0.994, demonstrating the effectiveness of hybrid feature engineering and intelligent sampling in DR detection.