Application Research Of Full Convolutional Neural Network In Fundus Images

The development of medical and computer technology has made it possible to analyze clinical data using computer technology.Fundus image is a commonly used medical experimental image,which can be used to help doctors to provide reference for the diagnosis of certain diseases,such as glaucoma,age-related macular degeneration,hypertension and diabetes.Among them,glaucoma is a highly dangerous and extensive eye disease that can cause permanent loss of vision.Cup-to-disk ratio is one of the important parameters for glaucoma screening,which requires precise segmentation of the cup and the optic disc.In addition,the macular area of the retina is one of the most important anatomical structures in the fundus image of the retina,an important part of the eye’s sensitization,and is located in the darker and pigmented areas of the retina.It is very necessary for its localization and segmentation are in the automatic analysis of retinal diseases.In this thesis,an experiment using a VGG16-based full convolutional network model is proposed,and a public DRIVE fundus image data set is used.In view of the small data samples in this data set,in this thesis,the method of image transformation is used to expand the experimental samples.In addition,the proposed model is further fine-tuned by borrowing the idea of the jump structure in the full convolution network,that is,the output model of the basic model and the second layer of the pooling layer is selected and merged,and the generated model is the primary micro-adjustment model.Then the model is superimposed with the output of the first layer of the pooling layer,and the resulting model is the ultimate micro-adjustment model.Secondly,using the idea of migration learning,the training parameters of the model are initialized,and the experimental data is trained and segmented by fine-tuning the parameters,and finally the segmentation result of the model is output.Then,the results of the segmentation of the primary micro-adjustment model and the ultimate micro-adjustment model are compared and analyzed.Through experimental comparison,it is found that by continuing to merge the feature map of the pooling layer,the segmentation effect of the three structures of optic cup,optic disc and macula can be improved to a certain extent.Secondly,the segmentation result is post-processed by the method of fully connected CRF,the cupsegmentation in the fundus image is greatly affected by the blood vessels.However,This method can be used to segment fundus blood vessels in subsequent projects.