DRG-NET: A graph neural network for computer-aided grading of diabetic retinopathy

• Published in: Signal, image and video processing Vol. 16; no. 7; pp. 1869 - 1875
• Main Authors: Salam, Amritha Abdul, Mahadevappa, Manjunatha, Das, Asha, Nair, Madhu S.
• Format: Journal Article
• Language: English
• Published: London Springer London 01.10.2022; Springer Nature B.V
• Subjects: Computer Imaging; Computer Science; Datasets; Diabetes; Diabetic retinopathy; Digital imaging; Digitization; Euclidean geometry; Feature extraction; Graph neural networks; Graphical representations; Image Processing and Computer Vision; Model accuracy; Multimedia Information Systems; Neural networks; Original Paper; Pattern Recognition and Graphics; Retinal images; Signal,Image and Speech Processing; Vision; Feature descriptor; Non-Euclidean geometry; Retinopathy; dimensional graph neural network
• ISSN: 1863-1703; 1863-1711
• DOI: 10.1007/s11760-022-02146-x

Diabetic retinopathy is emerging as a very serious vision disorder in the recent decades, due to escalation of diabetes world-over. This condition can be minimized to a great extend with timely prognosis. Computer-aided detection techniques are very useful for assisting ophthalmologists, for faster diagnosis and intervention. With the advent of digital fundus cameras and the digitization of retinal images, there is a huge availability of digital fundus images with expert-annotated labels. For addressing the challenge of digital image grading, an attempt was made to model the features in digital fundus images, utilizing the non-Euclidean geometry. Here, a Graph Neural Network with supervised learning is suitably adapted for diabetic retinopathy image grading. The images are represented as 3D graphs, to encapsulate discriminate information, as nodes in network. The features extracted from the diabetic retinopathy images, using Scale Invariant Feature Transform technique, is used for graph construction and training. The Diabetic Retinopathy Graph Neural Network namely, DRG-NET model is trained and validated on two publicly available datasets namely Aptos 2019 and Messidor. Ten different types of performance indicators, including accuracy and Cohen’s kappa values, were estimated and used for the comparison of models. For the Aptos and Messidor dataset, the model achieved an accuracy of 0.9954/0.9984, F1-score of 0.9774/0.9968 and kappa score of 0.9930/0.9980, respectively. It is evident from the results that the proposed DRG-NET model shows state-of-the-art performance for retinal image grading.