Automated Diabetic Retinopathy Detection Using Histogram of Oriented Gradients Features and Random Forest Classification: A Comparative Study with Deep Learning Approaches

Abstract

Diabetic retinopathy remains a leading cause of preventable blindness, necessitating efficient automated screening systems. This study investigates traditional machine learning techniques for diabetic retinopathy classification using the APTOS 2019 dataset. The methodology employs Histogram of Oriented Gradients feature extraction combined with Random Forest classification to categorize retinal fundus images into five severity levels. Experimental results demonstrate 94.00% overall accuracy with precision, recall, and F1-scores of 94.07%, 94.00%, and 94.00% respectively on 733 test images. Comparative analysis against a baseline Convolutional Neural Network reveals only 1.82 percentage point accuracy reduction while offering substantial computational efficiency advantages. The confusion matrix indicates 268 correctly classified diabetic retinopathy cases with balanced performance across severity classes. These findings demonstrate that carefully engineered traditional machine learning approaches achieve clinically relevant diagnostic accuracy suitable for resource-constrained healthcare settings, providing a computationally efficient alternative to deep learning methods for large-scale screening programs.