OHGCNet: Optimal feature selection-based hybrid graph convolutional network model for joint DR-DME classification

Abstract

Diabetic retinopathy (DR) is the crucial eye disease, which effects the blood vessels of the patient suffering with diabetes. The Diabetic macular edema (DME) is another crucial disease, that arises when DR reaches and damages the macula, resulting in fluid buildup in the retina. Individual, and joint screening methods of these DME, and DR require experts to manually analyze color eye fundus images. However, developing an efficient screening-oriented therapy is a time-consuming and costly endeavor because of the difficult nature of the screening approach and a scarcity of qualified human resources. In addition, automated systems are attempting to deal with these issues, and standard machine learning and deep learning processes have failed to fulfil the required criterion of performance and accuracy. Thus, this article focuses on the implementation of graph learning-based graph convolutional network (GCN) for the classification of joint DR-DME with enhanced accuracy. Initially, hybrid GCN (HGCN) with relation aware channel-spatial attention (RACSA) model is developed for extracting the deep features of individual DME, DR, and joint DR-DME. Further, a novel bio-optimization approach named modified deer hunting optimization algorithm (MDHOA) is employed as an optimal feature selection technique for the extraction of salient features. Finally, HGCN is utilized as a classifier for the classification of individual DME, DR, and joint DR-DME diseases. The extensive simulations conducted on IDRiD dataset shows that proposed OHGCNet performed superior as compared to the conventional methods with the improvement in classification accuracy as 5.11%, 3.88%, and 5.47% for DME, DR, and joint DR-DME. Furthermore, the performance of proposed OHGCNet also compared with the ISBI-sub challenge 2 and resulted in superior position as compared with leadership contenders.