Research On Detection Method Of Diabetic Retinopathy Based On Deep Learning

Deep learning algorithms can extract features from images and are widely used in the field of medical images.In severe cases,diabetic retinopathy can lead to blindness.At present,China’s medical resources are scarce,patients are difficult to see a doctor and other problems.The application of deep learning algorithm to the detection and analysis of characteristics of diabetic retinopathy can alleviate the above situation and bring good news to doctors and patients.The main research contents of this paper are as follows:The region of diabetic retinopathy was detected based on the improved Segnet network model.Segnet network model belongs to encoding and decoding structure.The convolution operation in the encoding stage is changed to VGGNET19,and the residual block is added in the process of convolution operation,which can effectively avoid gradient descent.In this experiment,Diaret DB data set is used for training and testing,and good results are obtained.The accuracy of the improved network was 0.955 for microaneurysms,0.853 for bleeding,0.714 for hard exudate and 0.898 for soft exudate,respectively.The experimental results show that this network can effectively detect diabetic retinal lesions.The classification of diabetic retinopathy grade was studied using an improved Googlenet network.As the depth of convolutional neural network deepens,the effect of feature extraction becomes better.The Googlenet network model has increased in depth and width compared to other networks,but at the same time it has increased the computational effort.In order to reduce the amount of calculation,in addition to adding dimensionality reduction operations to the Googlenet model itself,this paper replaces the 5×5 convolution kernels with two 3×3 convolution kernels in the Inception module to reduce the parameters without changing the effect.The research on classification of lesion grade is mainly conducted on the Messidor dataset for training and testing.The experimental results show that the improved Googlenet model can ensure the accuracy while reducing the calculation parameters,and can carry out the classification operation well.