Optimal hybrid feature selection technique for diabetic retinopathy grading using fundus images

Diabetic retinopathy (DR) has become the major cause of blindness for diabetic patients. This is because the microvascular consequence of diabetes mellitus results in DR, and treatment is successful only at the early stages. So, timely identification of DR is very important to minimize the risk of permanent vision loss. However, identifying and analyzing DR takes a long time and requires skilled ophthalmologists and radiologists. An automatic DR detection technique is needed in real-time applications to limit potential human errors. This paper proposes a hybrid bi-stage feature selection model for DR grading using the fundus images. Initially, the deep ensemble model extracts the efficient retinal features from preprocessed fundus images. Then, the proposed bi-stage feature selection method selects an optimal set of features to classify DR. In the first stage, two-filter-based feature selection techniques, namely Minimum Redundancy Maximum Relevance and Chi-squares, select the Guided features. In the second stage, the whale optimization algorithm reduces the feature space and selects more relevant and optimal features. The final optimal feature set is used for DR classification using support vector machines. The performance of the proposed model has been evaluated on the three publicly available databases, IDRiD, MESSIDOR-2, and Kaggle, and obtained an accuracy of 98.92%, a sensitivity of 99%, specificity of 99.69%, a precision of 98.8%, and F1-score of 0.988 with optimal features, which are better than other methods.