Optimization of LBP Texture Feature Extraction using Correlation And Mi For SVM-Based Diabetic Retinopathy Classification

Abstract

Diabetic retinopathy (DR) is a leading cause of blindness, making early detection based on retinal fundus images crucial. This study proposes a DR classification method with a primary contribution in feature optimization: integrating the LBP Contrast feature with a Local Binary Pattern (LBP) histogram and performing hybrid feature selection based on Mutual Information (MI) to assess relevance and correlation analysis to reduce redundancy. This method was tested using 168 images from the public Messidor dataset, with 100 images for training and 68 for testing to evaluate performance. Classification was performed using a Support Vector Machine (SVM) with a linear kernel, where model performance was evaluated before and after optimization to measure the significance of the improvement. The results showed a significant improvement after optimization, with accuracy increasing from 88% to 94%, recall increasing from 88% to 100%, and F1-score increasing from 0.92 to 0.96. Although precision decreased slightly from 96% to 93%, increasing recall to 100% is considered more crucial in a medical context as it minimizes the risk of missed positive cases. These findings confirm that the proposed feature optimization approach can significantly improve the accuracy and reliability of the DR detection system, offering potential clinical relevance for supporting early intervention.