Automatic Discrimination of Fundus DR Based on Improved Residual Dense Block Network

Diabetic retinopathy is the most serious complication of diabetes. In hospital treatment or telemedicine, experts analyze and treat patients with Diabetic Retinopathy (DR) based on the retinal images captured by the fundus camera. However, large number of non-pathological fundus images occupy too much time for the ophthalmologist to diagnose, and delay the timely treatment of patients with fundus DR. Therefore, it is a very urgent task to automatically and objectively screen whether the fundus has DR. Based on deep learning, we proposes an improved residual-dense module convolutional neural network structure (Modified Residual Dense Block Convolution Neural Network, MRDB-CNN). DR fundus images and non-DR fundus images are used for model training and the overall accuracy of the network structure is assessed by test set. Experiments have proved that the module can extract the detailed features of the fundus DR. The MRDB-CNN network structure can obtain a better generalization ability and a higher-precision network classification model while avoiding the complex image preprocessing. The accuracy of DR discrimination reached 94.90%, which reaches the needs of initial screening of fundus DR in hospital treatment and telemedicine.