Algorithm for diabetic retinal image analysis based on deep learning

In the last several years, diabetes incidence has increased year over year. Due to the large contrast between ophthalmologists' numbers and diabetic patients' numbers, many patients with diabetic retinal disease cannot be diagnosed and treated on time, worsening their condition. An improved Faster-RCNN network was designed to detect the location of the optic disc in the diabetic retina using object classification. In this study, the selection of the improved VGG16 model and ResNet50 model for feature extraction and the Faster-RCNN algorithm-based object classification detection for fundus images was experimentally investigated. The experimental data show that the ResNet50-based Faster-RCNN network model has higher average precision and faster model convergence in detecting diabetic retinal disease, achieving a mean average precision (mAP) of 97.42% and a precision of 98.96%. In the case of categories with more minor distinct features and smaller datasets, the precision is enhanced by 24.26% over the Yolov5-based object detection approach. The mAP value is improved by 6.12%. The algorithm better balances the precision and speed of object detection to better meet the requirements for detecting diabetic retinal disease.