Diabetic Retinopathy Detection Using DL-Based Feature Extraction and a Hybrid Attention-Based Stacking Ensemble

Diabetic retinopathy (DR) poses a significant threat to vision if left undetected and untreated. This paper addresses this challenge by utilizing advanced deep learning (DL) algorithms with established image processing techniques to enhance accuracy and efficiency in detection. Image processing extracts critical features from retinal images, acting as early warning signs for DR. Our proposed hybrid model combines image processing and machine learning (ML) strengths, leveraging discriminative abilities and custom features. The methodology involves data acquisition from a diverse dataset, data augmentation to enrich training data, and a multistep image processing pipeline. Feature extraction utilizes ResNet50, InceptionV3, and visual geometry group (VGG)-19 and combines their outputs for classification. Classification employs a decision tree (DT), K-nearest neighbor (KNN), support vector machine (SVM), and a modified convolutional neural network (CNN) with a spatial attention layer. Our work proposed a hybrid attention-based stacking ensemble with the mentioned models in the base layer and logistic regression model as meta layer, which further enhanced accuracy. The system, evaluated through metrics like confusion matrix, accuracy, and receiver operating characteristic (ROC) curve, promises improved diagnostic capabilities. The proposed methodology yields an accuracy of 99.768%.