Abstract: Objective  To evaluate the value of GhostNet lightweight network based on fundus color imaging in the diagnosis of diabetic retinopathy (DR), and the value of differential diagnosis between DR and other fundus diseases compared with classical convolutional neural network DenseNet121.Methods  Totally 29 535 color fundus photographs (including 9 883 DR, 2 000 normal fundus and 17 652 other blinding fundus diseases for differential diagnosis) were collected from large samples. Classical convolutional neural network (DenseNet121) and lightweight network (GhostNet) were used to train the model by transfer learning. Receiver operating characteristic (ROC)curve, sensitivity, specificity, positive predictive value, negative predictive value and area under the curve (AUC) were used to evaluate the performance of the two models. Results  Compared with model based on DenseNet121, the diagnostic time of single fundus photo by GhostNet model was shortened by 60.3%. In the diagnosis of DR, AUC value, sensitivity and specificity, accuracy reached 0.911, 0.888, 0.934, and 91.3% respectively. It is slightly lower than 0.954, 0.921, 0.986, and 95.5% of model based on DenseNet121. In the differential diagnosis of DR and other fundus diseases, the AUC value, sensitivity, specificity, accuracy of model based on GhostNet were 0.862, 0.856, 0.901, and 87.8% respectively. For model based on DenseNet121, the values were 0.899, 0.871, 0.935, and 90.2%, respectively. Conclusions  The operation speed of DR diagnosis model and differential diagnosis model based on GhostNet lightweight neural network is significantly faster than that of classical model DenseNet121, and the accuracy is satisfying. For primary medical institutions with lack of ophthalmologists and low equipment performance, such as community hospitals, we can consider using this technology to carry out primary screening of DR.

Key words: diabetic retinopathy, lightweight neural network model, diagnosis, screening, community