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

Purpose: 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. Methods: 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. Results: The architectures were assessed using (1) the performance metrics - accuracy, precision, recall, and F1score, (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. Conclusion: 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 Logistic Regression. Another best-performing architecture is similar with a variation in DenseNet121 for feature extraction and support vector as a meta-learner.