Research On The Detection Of Exudates In Diabetic Retinal Images Based On Deep Learning

Detection of diabetic retinopathy is to use medical image processing technology to detect lesions in the fundus images,and then mark the location of the lesion.By using the deep learning,the intelligent diagnosis and analysis of the hard exudates in the diabetic retinopathy provide the basis for the early diagnosis of diabetes,which can also reduce the cost of medical treatment,share the pressure of doctors and improve the accuracy of the diagnosis.Based on the analysis of the relevant work at home and abroad,the detection methods of hard exudates based on principal component analysis network(PCANet)and U-Net are realized and tested in the e-ophtha EX public database in the pixel level.The main works of this article are as follows:1)A method to extract the candidate regions of hard exudates is realized.Because the hard exudates and the optic disc have a certain similarity,it is necessary to exclude the position of the fundus background,the main blood vessel and the optic disc with the priori knowledge.So the main blood vessel is firstly extracted.Then the optic disc is located according to the projection of the main vessel,and the candidate image of the hard exudates is finally obtained.2)After obtaining the candidate regions of hard exudates,a method for detecting hard exudates based on PCANet is realized.PCANet is expanded from the recognition field to detection and segmentation field,which achieves to detect hard exudates by using principal component analysis to map the detection of object pixel and adjacent pixels to the predicted probability.The problem of unbalance of sample distribution caused by the small proportion of hard exudation in the image is solved by random sampling based on labeled image,and the problem of dimension explosion in the original PCANet is improved by using compressed sensing.At the same time,multi-scale convolution kernels and multi-scale features are used to describe accurate local features,which makes the new network complete the detection task better.In addition,in order to avoid repeating doing convolution in the adjacent blocks,the sampling of the input layer is delayed to the output layer in testing process.In addition,by sampling the image blocks of hard exudates with the candidate region,the speed and accuracy of the detection are improved according to these prior knowledge.At the same time,the sampling is changed to the input layer to avoid the repeated convolution operation.3)After obtaining the candidate areas of hard exudates,a method of detecting hard exudates based on U-Net is implemented.In view of the limited scale of samples,the network is used to study the image blocks of hard exudates and achieve detection from end to end.In order to solve the unbalanced problem in training samples,it is improved by random sampling based on labeled images and selecting appropriate cross entropy loss function.Then,it is tried to use the multiply convolution structure to extract multi-scale features structurally,which is convenient to make the network more wide and deep.In addition,during the convolution process,the edge of the image block is filled with the nearby pixel to align the image.This leads to the error detection at the edge of the image.So the pixels in the result image take the mean value of predicted value for themselves,which improve the detection.The experiments show that the method is more accurate than the traditional method,and the detection of hard exudates can reach a high level.