Research On Intelligent Diagnosis Technology Of Diabetic Retinopathy

Diabetic retinopathy(DR)is a major cause of visual impairment and blindness in humans.Clinical studies show that only early DR screening and timely diagnosis of diabetic patients can prevent their vision decline.At present,the disease is mainly treated by manual diagnosis.However,with the increase of the number of screening,it is very easy for doctors to miss diagnosis,misdiagnosis,feedback is not timely and other problems,so as to miss the best treatment opportunity of the disease.Therefore,it is urgent to develop effective intelligent technology for disease diagnosis.In recent years,intelligent technologies such as machine learning and deep learning have been widely applied in the diagnosis of diabetic retinopathy,playing an important role in assisting doctors in large-scale screening and improving classification accuracy and diagnostic efficiency.Combined with classical image processing technology,this paper studies DR classification based on machine learning and deep learning methods.The specific research contents are as follows:(1)In order to solve the interference caused by uneven image illumination and unbalanced data,pre-processing operations such as clipping,grayscale,color and brightness normalization,image enhancement,gamma correction and image expansion are adopted.The image expansion adopts flip,translation,rotation and so on.(2)Based on the problem that existing models only fuse depth features,a multi-category feature fusion classification model is proposed.According to the characteristics of fundus lesions,on the basis of extracting Alex Net,VGG-16 and RESNET-50 depth features,four new image omics features,including Gabor,LBP,HOG and Haralick,are added,and the four image omics features and three depth features are screened by the two-layer feature fusion network.Finally,SVM,RF,k NN and ELM are used to classify the lesions,and the accuracy reaches 88.64%,86.63%,86.27% and 85.61%,respectively.It effectively avoids the problem of feature loss caused by insufficient feature extraction at the bottom of deep network and less feature dimension of image omics.(3)In order to improve the model’s attention to small lesions,a classification model based on depth separable convolution and attention mechanism is proposed.The attention mechanism is based on 1×1 convolution and activation function to identify key features,and the lightweight module is used to replace the standard convolution of the model,so as to achieve rapid classification of lesions while ensuring accuracy.After verification on the data set,its classification accuracy reaches 91.5%,recall rate reaches 90.3%,parameter number is 12.24 m,FLOPs is 9.74 g,running time is 158.7ms,Kappa value is 0.893.In addition to solving the problem of excessive irrelevant information such as background and structure of multi-category feature network extraction,the model parameters are effectively reduced and the running time is reduced while the attention to the features of small lesions is enhanced.