The Research On Fundus Retinal Image Segmentation Based On Deep Learning

Diabetic Retinopathy(DR)is the leading cause of blindness in adults.At present,DR diagnosis is usually diagnosed by observing abnormal retinal lesions by ophthalmologists.However,DR has a great variety of lesions with varying degrees,leading ophthalmologists difficult diagnosis.Therefore,the use of computer vision technology can assist ophthalmologists to achieve diagnosis,not only can improve the diagnostic efficiency,but may also increase the diagnostic accuracy.In this paper,the Convolutional Neural Network(CNN)is used to segment fundus images.Aiming at the problem of small fundus image data set,two solutions are proposed:1.Use entropy sampling method to extract image patches effectively and design patchbased fully convolutional networks.Entropy sampling can eliminate a large number of patches that do not contain valid information,greatly reducing the amount of computation and taking full advantage of the information of neighboring pixels to improve the segmentation performance.To solve the category imbalance problem,weighted cross entropy is used as a loss function.2.Data augmentation is used to augment the dataset,which can be used to train Fully Convolutional Network(FCN)for image segmentation.The standard U-Net uses long skip connections for efficient delivery of features for better segmentation,while Res Net shortcuts and Dense Net dense connections are short skip connections.Through enhancing the gradient flowing in a local area,the network can be trained much more easier.To take full advantage of the short and long skip connections,we introduce the residual structure and the dense structure into the standard U-Net,which are called residual UNet and dense U-Net respectively.Finally,we use the extended Dice score as the loss function.In this paper,the segmentation of fundus images include blood vessel segmentation,optic disc segmentation,exudate segmentation and hemorrhage segmentation in four parts.We segmented blood vessels using a patch-based fully convoluted network with segmentation results reaching 78.11% sensitivity at the pixel level,98.39% specificity,95.60% accuracy,87.36%Precision and 97.92% AUC score.The proposed method has achieved good experimental results.At the same time,the residual U-Net and dense U-Net are used for segmenting the optic disc,exudate and hemorrhages in retinal image,which have achieved good experimental results.The segmentation algorithm proposed in this paper has been tested in Shanghai First People's Hospital.Finally,by visualizing some feature maps,we can find that the initial convolution layer is mainly used to extract low-level features including lines,corners and edges.As the network deeper,the convolution layer mainly extracts the advanced semantics Features,which contains no details.In the segmentation problem,upsampling layer must be applied.We can find more and more details in features after upsampling layer is applied.Through the visualization of feature maps,it is easy to find out the reasons that may affect the performance of segmentation and provide ideas for further improvement.