PERFORMANCE COMPARISON OF DIABETIC RETINOPATHY CLASSIFICATION FROM RETINAL IMAGES USING TEXTURE FEATURES AND MACHINE LEARNING

Abstract

This research aims to classify diabetic retinopathy in fundus images using machine learning combined with texture feature extraction and to compare the performance of various machine learning models for diabetic retinopathy classification. The study employed a dataset of 8,834 images published on www.kaggle.com, collected by Chaudhari (2020). According to the study, texture features were extracted from fundus images and converted into numerical statistical features. Redundant features were reduced using Spearman's rank correlation coefficient with a correlation threshold at 0.90 to improve data quality. The number of irrelevant features were further reduced using the ANOVA F-test by selecting the top 100 features with the highest F-scores showing most strongly correlated with the classification outcomes. Subsequently, feature importance analysis using Random Forest were applied to select statistical features with importance values ≥ 0.01, which were considered significant features for model development. This entire process resulted in 34 features selected from the original 316 features. To compare model performance, five machine learning models were trained using these selected features: Random Forest, eXtreme Gradient Boosting, Support Vector Machine, Neural Network, and K-Nearest Neighbors. The results demonstrate that the Extreme Gradient Boosting model achieved the highest performance with 99.34% accuracy, 99.35% overall performance, 99.98% sensitivity, and 98.71% specificity. This approach exhibited superior classification performance with enhanced interpretability compared to modern deep learning methods. Despite utilizing a smaller dataset than typically required by deep learning approaches, it maintained competitive classification performance while being particularly suitable for deployment in resource-constrained clinical environments.