Recent years, deep learning in the image identification has made great progress, showing good application prospects in medical image reading. Diabetic Retinopathy (DR) is an eye disease due to diabetes, which is the most ordinary cause of blindness. Traditional diabetic retinopathy detection is a manual and time-consuming and labor-intensive process, which requires a highly experienced clinician to examine and evaluate the digital color fundus photos of the retina. Therefore, it is crucial to use the deep learning technique to achieve automatic detection of diabetic retinopathy. In this paper, we proposed a diabetic retinopathy detection method based on deep learning and proposed a network structure named multi-self-attention. At first, the image features were extracted through the InceptionV3 model, and then the feature maps was directly generated. Secondly, the feature maps, which can reflect condition of retina, will be input into multi-self-attention network structure, to calculate multi-self-attention feature. Finally, by convolutional layer and fully connected layer, the stage results about diabetic retinopathy will be obtained. With the experiments in TensorFlow framework , the effectiveness of multi-self-attention network structure for feature extraction and classification is proved.