Application of End-To-End EfficientNetV2 in Diabetic Retinopathy Grading

Diabetic macular edema is a common cause of vision loss in patients with diabetic retinopathy, there are currently about 21 million diabetic macular edema patients worldwide. In this paper, we propose an automatic diagnosis method of diabetic retinopathy grading based on deep learning. This makes the disease diagnosis process more convenient and efficient. We preprocess the image including dark channel-based dehazing and color image histogram equalization to enhance the contrast and image brightness of the image. We propose a convolutional neural network called "RTNet" based on the EfficientNetV2 net-work. On the basis of the original squeeze-and-excite (SE) attention mechanism, two pooling layers are added, which can effectively reduce the size of the image, improve the calculation speed, effectively reduce the dimension and reduce the amount of calculation. Since the Stochastic Gradient Descent (SGD) optimizer learning rate adjustment strategy is limited by the pre-specified adjustment rules, we adjust its SGD optimizer to Adadelta optimizer, finding the parameters can be moved closer to the bottom of the slope, in order to speed up the convergence. The recognition accuracy rate reached 84.2%, which proved that RTNet can accurately identify diabetic retinopathy and has great application prospects.