Research Of Diabetic Retinal Image Analysis Algorithms Based On Deep Learning

The study of artificial intelligence has received much more attention from the whole community in recent years and there is no doubt that deep learning is very popular in this field.Combining artificial intelligence and medical technology,artificial intelligence medical technology provides the basis for disease diagnosing under the help of computer technology,which is much more useful in the modern information society.The retinal fundus images can reflect much information about some diseases.Traditional processing methods for the images are in stages,which is not only excessively dependent on doctors'experience,but also time-consuming and inefficient.Therefore,a more intelligent retinal image analysis system is very necessary.This paper presents a complete automatic identification system for retinal images based on deep neural network.Considering the particularity of retinal dataset,the image preprocessing for diabetes retinal images is primary.After that,CompactNet,a compact deep neural network model,is proposed to apply to the system.In the proposed method,CompactNet inherits the shallow structure of AlexNet and while parameters of deep network are adjusted adaptively according to the training data.Finally,to validate the effectiveness of the network model,experiments are conducted on different training methods with various network structures.The experimental results demonstrate that the fine-tuning method of CompactNet outperforms the traditional training methods and the evaluating indicator can reach 0.87,increased by 0.25.In addition,compared with LeNet and AlexNet,CompactNet obtain the highest classification accuracy.Experiments also prove that it is necessary to preprocess the data,such as data augmentation.The segmentation of blood vessels has always been the mainstream approach for analyzing retinal images.Restricted by the specificity of retinal images,the traditional segmentation methods have a poor generalization ability.In this paper,the deep learning method is used to achieve the point-to-point segmentation of the retinal fundus image.The input image can be directly used as a training sample.Vascular binary map can be directly obtained by post-processing synthesis of output.RetinalSegNet is available for input images of any size without dimension limitation.Experiments show that the results of proposed method achieve a better continuity and a generalization performance.