Research On The Segmentation Method Of Targets In Fundus Images Based On U--shaped Neural Network

The use of deep-learning to process medical images has gradually become a hot issue with the rapidly development of deep neural networks technology.Fundus images is a type of common medical image in eye disease screening,it mainly includes the optic disc,optic cup,blood vessels and the macular area which can provide important diagnostic basis for certain diseases.For instance,the diameter ratio of optic cup to optic disc can be used for the diagnosis of glaucoma,hemorrhage,exudation,microvascular abnormalities,etc.can be used as a basis for the diagnosis of diabetes,retinal arteriopathy may be one of the symptoms of hypertension.In order to accurately interpret the fundus images and make corresponding medical evaluation,it is necessary to accurately segment the optic disc,optic cup and blood vessels in the fundus images.However,manual interpretation of fundus images relies heavily on physicians’ experience and is costly in terms of time and labor.For these reasons,it is important to segment valuable structures in fundus images with high accuracy by means of deep learning to diagnose fundus diseases.In this thesis,we investigate the segmentation methods of optic disc,optic cup and blood vessels in fundus images based on domestic and international studies.The main research includes 3points.In this thesis,we first carried out the research on the segmentation method of optic disc and optic cup,and proposed a pooling module based on attention mechanism for the problem of image information loss in existing segmentation methods,and then designed a fundus image segmentation network-APUNet for optic disc and optic cup segmentation based on this module combined with U-shaped neural network structure.Finally,we con-ducted the experimental results prove that the APUNet proposed in this thesis can accu-rately fit the contours of optic disc and optic cup and generate the segmentation prediction result images of optic disc and optic cup.Then,we explores the vessel segmentation method based on the optic disc and optic cup segmentation method,and proposes a vessel segmentation prediction process based on image cropping for the problem that the existing segmentation methods cannot accu-rately segment the vessel endings,and performs experimental validation to generate the vessel segmentation prediction result images.Through the comparison of experimental results,it was verified that the fundus image segmentation method proposed in this thesis could segment the fundus vascular structures more accurately and the segmentation effect surpassed the existing segmentation methods.Finally,we also investigates the optimization method of network training based on transfer learning,and proposes three different transfer strategies for the problems of slow convergence and unstable model training in the training of fundus image segmentation network,and conducts experimental validation to accelerate the convergence speed of fundus image segmentation network and further improve the accuracy of segmentation.