Diabetic Retinopathy Classification Using a Hybrid Deep Learning and Machine Learning Model

Abstract

Objective: Among diabetic patients, diabetic retinopathy (DR) remains one of the most common causes ofpreventable blindness and vision loss, making its early detection crucial for preventing irreversiblecomplications. Manual evaluation of fundus photographs is a lengthy process. Additionally, it requiresspecialized training that is not always available in all clinical settings. Consequently, artificial intelligence-basedautomated retinal image analysis systems have emerged as complementary tools to enhance diagnostic accuracyand efficiency. This study proposes an ensemble learning-based framework to improve the accuracy androbustness of automated DR detection. In the first stage, pretrained convolutional neural network (CNN) modelsextract high-level features from fundus images, capturing complex patterns and DR-related lesions. Thesefeatures are then fed into several classical machine-learning classifiers, including Support Vector Machine(SVM), Random Forest, and XGBoost. To further boost discriminative power and reduce classification errors,a stacking ensemble strategy integrates the predictions of the individual classifiers within a meta-learningframework, enabling the model to learn the optimal combination for DR detection and grading. This hybridapproach effectively combines the strengths of deep learning and classical machine learning, yielding improvedperformance in DR detection and classification. Experimental results show that the stacking ensemble achieveshigher accuracy and F1-score compared to individual models, underscoring its potential as an auxiliary tool for early diabetic retinopathy detection.