A hybrid approach with customized machine learning classifiers and multiple feature extractors for enhancing diabetic retinopathy detection

Abstract

Diabetic retinopathy (DR) is a severe global issue causing blindness if untreated, affecting millions worldwide and worsening over time. Addressing this growing concern necessitates early and accurate DR identification. This study introduces a novel approach to DR detection, combining machine learning algorithms with deep feature extraction techniques. A hybrid model is proposed by stacking predictions from diverse classifiers, such as Decision Trees, Random Forests, Support Vector Machines (SVMs), and more. Three deep learning models – MobileNetV2, DenseNet121, and InceptionResNetV2 – are employed as feature extractors from retinal images. Each classifier undergoes hyperparameter tuning for optimal performance. Using the APTOS 2019 Blindness Detection dataset, including preprocessing techniques like data augmentation and standardization, this hybrid model demonstrates promising accuracy in multi-class (95.50%) and binary classification (98.36%). Notably, DenseNet121 outperforms others. The results suggest the effectiveness of this hybrid technique in early diabetic retinopathy detection, holding significant promise for improved medical intervention.

Keywords

• Machine learning
• Diabetic retinopathy
• Hybrid model
• Synthetic minority over-sampling technique (SMOTE)
• Stacking