Research On Vessel Segmentation And Detection Of Diabetic Retinopathy In Color Fundus Images Based On Convolutional Neural Network

The eye is the window of the soul and one of the key organs for the human body to perceive the external environment,and the retina is the core component of the eye organ,and its health status directly affects the human body's vision level.Among the existing patients with retinal damage,retinopathy caused by diabetes accounts for a considerable part,especially in the middle-aged and elderly population,which has become the leading cause of visual impairment.The reaction more or less came out,and the fundus retina image can directly observe the health state of the capillaries,and discover the abnormal state of the body in time.Therefore,relevant research on retinal fundus images can not only provide help for the treatment of diabetic patients with retinopathy,but also provide a scientific basis for the early detection and treatment of other related diseases.This thesis takes color retinal fundus images as the research object,conducts research on vascular image segmentation method and retinopathy detection theory,based on the existing medical image processing methods,introduces convolutional neural network analysis theory,tries to design and build an image network analysis model with better applicability can effectively improve the accuracy of blood vessel segmentation based on fundus images and the detection efficiency of diabetic retinopathy.The specific research results and contents are summarized as follows:1?An ensemble residual U-shaped segmentation network(ERU-Net)is proposedIn view of the limitations of current color retinal fundus image blood vessel segmentation methods,such as low precision and slow speed,this paper designs an integrated residual U-shaped segmentation network,which has achieved good results in improving the accuracy and speed of fundus image segmentation.First,a simple residual module is designed to solve the problem of information loss and network training slow when sampling in the U-Net structure;then,a kind of residual module and sampling module is designed.U-shaped basic network;then,using the ensemble strategy,the four basic networks of the same type are assembled into a large segmentation network,as far as possible to improve the accuracy of fundus image segmentation;finally,the experiment is used to verify the design of the network model practicability and effectiveness.2?A detection of diabetic retinopathy based on destruction and construction learning(DCL-Net)is proposedBased on the detection problems of diabetic retinopathy,this paper uses a destruction-construction structure to solve the problem of diabetic retinopathy detection and classification.Destruction-construction mechanism is to disrupt the global structure,retain local details,and force the recognition network to focus on distinguishing local regions for recognition,and then rebuild to restore the original regional layout.Based on this structural network,it can greatly improve the fineness of the image,and thus improve the performance of image recognition.Finally,through experimental comparison and verification,the design network has achieved good results in the detection of diabetic retinopathy under the graphical conditions of very little data and unbalanced data distribution.