Automatic Screening of Diabetic Retinopathy Based on Residual Network

Diabetic retinopathy(DR) is one of the major causes of blindness. It can cause visual impairment and even blindness. Timely treatment can save the patient’s vision as much as possible, so the screening of DR is very important. In order to solve the problem of low accuracy for DR screening, a new residual network based method using multi-class training data for DR screening is proposed. Firstly, by collecting, labeling and arranging fundus images, a new dataset for DR screening-MultiClassDR dataset, which contains three classes: health, DR and other fundus diseases is constructed. Secondly, the residual network model is constructed for the high-resolution image datasets and the model is pre-trained by ImageNet dataset and diabetic retinopathy detection dataset of Kaggle to obtain the basic features of the fundus images. Finally, the model is trained and tested on the MultiClassDR dataset. The average accuracy of the model for DR screening is 87.2%, indicating this method can improve the performance of the model for DR screening while improving the learning ability.