Computer-Aided Diagnosis-Based Grading Classification of Diabetic Retinopathy Using Deep Graph Correlation Network with IRF

Diabetic retinopathy (DR) is a leading cause of blindness in different age groups. So, the early detection of DR can save millions of people from blindness issues. Further, the manual analysis of DR requires much processing time and experienced doctors. So, computer-aided diagnosis (CAD)-based artificial intelligence models were developed for early DR prediction. These traditional methods, however, needed to extract balanced deep features, resulting in poor classification performance. This study creates a CAD-based grading classification of DR using graph learning and optimal feature selection properties, which is named as CEFNet. The EyePACS and Messidor datasets are first put through the edited nearest neighbor oversampling (ENNOS) method, ensuring that each class's instances are equal. Then, a deep graph correlation network (DGCN) is used to get the unique features of each class by finding the connections in color eye fundus (CEF) images. In addition, an iterative random forest (IRF) is performed for feature selection to select the highly relevant DGCN features. Finally, these balanced optimal features are used to train a decision tree-based extreme gradient boosting (DTEGB) classifier, which then uses test CEF image data to make the prediction. The simulations done on the EyePACS and Messidor datasets show that the proposed CEFNet performed better than the current best DR grading classification methods. The EyePACS dataset had an accuracy of 99.50%, a sensitivity of 99.20%, and a specificity of 100%. The Messidor dataset had an accuracy of 99.80%, a sensitivity of 99.67%, and a specificity of 100%.