Research On Segmentation Method Of Fundus Retinal Vessels Based On Deep Learning

In medical clinical diagnosis,there are a certain number of capillaries in the retina of the fundus,and the changes in their characterization are closely related to the concurrent symptoms of many diseases.The analysis and study of the distribution and morphology of these vessels is one of the necessary bases for the diagnosis of some ophthalmic diseases and comprehensive diseases.The analysis and study of the distribution and morphology of these vessels is one of the important bases for the diagnosis of some ocular diseases and comprehensive diseases.The accurate segmentation of retinal vascular images by computer can greatly reduce the workload of clinicians,improve the efficiency of the segmentation task and avoid the influence of human and subjective factors,which has far-reaching significance for medical aided diagnosis.In this paper,the segmentation algorithm of fundus retinal blood vessels based on deep learning algorithm is studied,two different network models are designed,and a large number of experiments are carried out in the public data sets Drive and STARE.The experimental results show that both of the two network models can accomplish the segmentation task effectively.The main work of this paper includes:(1)Aiming at the segmentation task of fundus blood vessels,this paper discusses the principle and application method of full convolutional network in deep learning algorithm.According to performance of pixel-level precise semantic segmentation of full convolutional network,the U-NET network model is studied and improved.(2)In order to improve the global receptive field of the full convolutional network and increase the segmentation accuracy of fine blood vessels,a retinal blood vessel segmentation model,IPDU-NET,based on multi-scale feature fusion,was proposed in this paper.This network adopts a symmetric codec network structure.Increased feature mobility in the network is achieved by using the Inception multi-scale feature extraction module instead of a single convolution layer in the original network.Pyramid void convolution is used in the pooling layer to further extract the context information.The experimental results show that the improved IPDU-NET has strong ability and accuracy of vascular segmentation.(3)In order to further improve the accuracy of various indicators of the level of fundus retinal vascular segmentation and the segmentation of small blood vessels,this paper proposed ATTR2U-NET based on U-NET network.This network integrates residual learning,secondary convolution and attention mechanism,and fuses residual learning structure and secondary convolution into R2CU module.In the process of up and down sampling,the image features are extracted deeply,the receptive field is increased,and the gradient problem is effectively prevented with the increase of network depth.By introducing attention gate AGS into jump connection and combining coarse and fine filters,the fine-grained extraction of retinal blood vessels by the training network model was increased,and the anti-noise ability of the model was improved.Experimental results show that this network model has good capability in segmentation.which can solve the problems of insufficient accuracy and sensitivity,and the segmentation fracture at the small vessel bifurcation.