Improving cataract diagnosis using ensemble learning and modified random forest (MRF) classifier

This study explores modern techniques to enhance cataract detection accuracy, including ensemble learning, hybrid feature extraction, and a Modified Random Forest classifier. Traditional methods face limitations in feature extraction and classification, which are addressed through a hybrid approach combining LBP, GLCM, and CNN. LBP detects early-stage cataracts, GLCM captures spatial relationships, and CNN extracts deep structural features, improving image representation. The proposed Modified Random Forest (MRF) classifier integrates feature weighting and optimized decision thresholds, reducing noise and enhancing classification accuracy. Feature selection with Recursive Feature Elimination (RFE) and Principal Component Analysis (PCA) minimizes overfitting and computational cost. Ensemble learning methods such as Bagging, Boosting, and Stacking further improve model robustness, with Stacking achieving 93% accuracy and high ROC-AUC for early-stage detection. However, computational complexity remains a challenge, particularly for deployment in clinical settings. KNN and SVM models underperform without feature selection, highlighting the need for careful preprocessing. Despite these challenges, the proposed techniques significantly improve cataract detection, ensuring better generalization across datasets.