Grading of Diabetic Retinopathy Images Based on Graph Neural Network

Diabetic Retinopathy (DR) has become one of the main reasons for the rise in the number of limited vision people worldwide, while high-definition color fundus images have brought great convenience to the diagnosis of DR. However, manual image reading is time-consuming and labor-intensive, and different doctors may make different diagnoses. At present, intelligent grading based on deep learning has become a hotspot in DR intelligent diagnosis. The existing DR intelligent classification model based on convolutional neural network has achieved good results, but the relationship between the deep features proposed by the network is not considered, while this relationship contains important classification information. In order to overcome the above-mentioned shortcoming of convolutional networks, this article draws on the powerful relationship capture capabilities of graph neural networks and proposes a new DR intelligent classification model. The model is composed of two cascaded networks. The convolutional neural network is used to extract the deep features of the DR image, and the graph neural network is used to further capture the relationship between the deep features of the convolutional network. Finally, the two network outputs are fused by adaptive weight, and give the grading result of the entire network. The proposed model is evaluated on the APTOS2019 and Messidor-2 datasets. Compared with other models, the grading accuracy and F1-score of the proposed model on APTOS2019 are improved by 1.1% and 1.3%, respectively. The grading accuracy and F1-score are improved by 1.4% and 1.8% on Messidor-2, respectively.

Automated summary

The article introduces a new intelligent classification model for diabetic retinopathy by combining convolutional neural networks and graph neural networks, which successfully improves the accuracy of diabetic retinopathy diagnosis.