Deep hybrid architecture with stacked ensemble learning for binary classification of retinal disease

•A deep hybrid architecture combining feature extraction using CNN with an ensemble of machine learning classifiers ensembled with the stacking method is developed.•The study evaluates 144 combinations to identify the most compelling end-to-end deep hybrid architecture for binary retinal disease classification.•The proposed architecture is trained and tested on the RFMiD dataset.•Each combination's performance is assessed using performance metrics: accuracy, precision, recall, and F1 score.•The robustness of the results obtained from this study is evaluated using various statistical tests and statistical graphic techniques.•The empirical analysis recommends deep hybrid architectures that achieved the highest accuracy of 92.34 percent on the RFMiD dataset. Early retinal disease identification is vital since symptoms are passive at initial stages but lead to irreversible vision loss at advanced stages. Globally, a substantial population is at risk of vision impairment, prompting researchers to investigate methods for efficient classification. This work experimented one hundred and forty-four different hybrid architectures amalgamating each of the eight convolutional neural architectures (VGG, EfficientNet, Inception, ResNet, NasNet, DenseNet, InceptionResNet, Xception) with seven classifiers (Logistic regression, K-Nearest Neighbours, Support Vector Classifier, Decision Tree, Bagging classifier, Random Forest, Adaptive Boosting, Light Gradient Boost and Extra tree classifier). The top performing n (n=3,4,5) classifiers were ensembled with meta-learner using stacking strategy. The performance of the pipeline is evaluated with two distinct meta-learners, three different image sizes, and feature counts on the RFMiD dataset. The architectures were assessed using (1) the performance metrics – accuracy, precision, recall, and F1-score, (2) statistical graphics to understand the prevalence of classifiers, Borda count voting method to identify the best CNN model, (3) Tukey's honestly significance difference test to identify best-performing architecture. Two architectures achieved a high accuracy of 92.34 percent and F1 scores of 95.11 and 95.19. This is the first work to experiment with 144 combinations to identify suitable deep architecture for binary retinal disease classification. The study recommends Xception for feature extraction ensembled with ExtraTreeClassifier, Light gradient boosting machine, Random Forest, AdaBoost classifiers, and meta-learner as Logist