DIABETIC RETINOPATHY CLASSIFICATION USING CONVOLUTIONAL NEURAL NETWORK AND STACKING ENSEMBLE OF CLASSIFIERS TECHNIQUES

Abstract

Early detection of diabetic retinopathy (DR) is critical, as prompt treatment can help reduce or even prevent visual loss. Most of the current state-of-the-art machine learning techniques for DR detection and classification make use of single classification for prediction. However this single classification models suffer from high variance, high bias, bottleneck in local optima, and the researcher also suffers the risk of choosing the wrong classifier. These issues can be solved by combining the predictions from multiple classifiers which produces predictions that are less sensitive to the specifics of the training data, the choice of training scheme and the serendipity of a single training run. Therefore, this research proposes an effective stacking ensemble technique for DR classification that will satisfy the drawbacks of using a single model, hence improve classification performance. The proposed stacking ensemble classifier was produced from a combination of four classifiers namely: Support Vector Machine (SVM), KNearest Neighour (KNN), Decision Tree (DT) and Naïve Bayes (NB). The proposed stacking ensemble technique was evaluated using the Messidor dataset. In comparison with the performance of the individual constituent models the proposed stacking ensemble technique achieved an acuracy of 99.17% which was better than the values achieved by the constituent models with accuracies of 98.33%, 96.67%, 93.33%, 95.00% for SVM, KNN, DT and NB, respectively. The proposed technique also produced better results than previous works based on Messidor dataset. These results suggest the robustness of the proposed model to DR detection and classification.