Research On Diabetic Retinopathy Detection Based On Convolution Neural Network

Diabetic Retinopathy(DR)is a chronic and imperceptible disease,which is one of the main causes of blindness.Therefore,it is necessary to screen DR early in the normal population.Large scale screening of fundus images increases the burden of doctors.In order to reduce the workload and improve the efficiency of doctors,it is of great significance to realize the automatic identification of lesions in DR screening system.Detecting the early lesions from the fundus images accurately is one of the key problems in large-scale DR screening system.It is particularly important to detect microaneurysms and exudates in the research of early detection of DR.However,the anatomical structure and lesions in the fundus image being similar in some characteristics and the poor quality of the fundus image caused by the imaging hardware conditions make it difficult to detect DR.The detection methods of DR lesions generally include three steps:preprocessing,extraction of candidate regions and identification of lesions(including the feature extraction and classification of lesions).The expression of the lesions is one of the key problems of the algorithm.Features of the lesions are mainly based on manual design,which not only requires a prior knowledge of the fundus image but also has a very complicated process.The extracted features are difficult to fully reflect the characteristics of the lesions,so the performance of the detection algorithm needs to be improved.Aiming at the problem of extracting pathological features and classification,the method of DR detection based on convolutional neural network(CNN)is proposed in this paper,the main research work is as follows:(1)Detection of microaneurysms.The color features of the microaneurysms are similar to those of blood vessels and hemorrhage and the edge is fuzzy,so a method of detecting the microaneurysms based on CNN and SVM is proposed in this paper.Firstly,the contrast limited adaptive histogram equalization is used to enhance the contrast of the microaneurysm,and the local minimum detection method is used to extract the candidate region of the microaneurysm.Secondly,we use the cross validation to train the CNN model of the microaneurysms.The network structure introduces batch normalization layer and dropout layer to prevent the over fitting of the model and accelerate the convergence speed.Then,the feature vectors extracted by the optimal CNN model are used to train the SVM classifier.Finally,SVM was used to detect the microaneurysms in candidate region.The method is tested on ROC and e-ophtha-MA data sets.The results show that the proposed method has good generalization ability.(2)Detection of exudates.The shape,size and location of exudates are different in fundus images.In this paper,a method for the detection of exudates based on CNN is proposed.Firstly,combining the HSV model and the mathematical morphology reconstruction method to extract the candidate region of exudates.Then,the cross validation is used to train the CNN model.Different from the classification model based on region,the detection of exudates is based on the pixel classification model.We build the different network structure.The batch normalization and dropout methods are also used to train and optimize the model.Finally,the optimal CNN model is used to classify the pixels in candidate regions.The method is tested on e-ophtha-EX and other data sets,and the good detection effect is obtained.