| The eye is a vital part of the human body.Its health is inextricably linked to human life.Doctors rely heavily on the interpretation of fundus retinal pictures when diagnosing eye disorders and various organ conditions.The essential condition for doctors to diagnose patients diseases is fundus vascular segmentation,which is of great significance for clinical medicine.As a result,it is necessary to study segment fundus retinal vessels for doctors to diagnose diseases effectively.Many methods based on deep learning have been applied to fundus retinal vascular segmentation in recent years,thanks to the rapid growth of deep learning in the field of computer vision.However,the contrast between blood vessels and background may be low in practice due to retinal vascular disease,shooting light,and other factors,resulting in poor recognition of fine blood vessels,under segmentation of blood vessels in complex curvature forms over segmentation,and low segmentation sensitivity.This research provides two distinct vascular segmentation techniques for fundus retinal images based on deep learning.The main research contents of this paper are as follows:(1)A multi feature fusion retinal vessel segmentation model based on codes structure is designed.Aiming at the problem of low accuracy caused by the loss or fracture of vascular details in vascular segmentation of fundus retinal image.A new retinal vessel segmentation algorithm is proposed based on the U-shaped network of encoding and decoding structure.The model completes information transfer better by introducing residual module and short hop connection to fuse high and low features.In addition,the model introduces void convolution and attention mechanism to expand the receptive field,enhance feature information and improve the generalization ability of the network.The overall performance of the network model has been improved by combining these improved modules cleverly.(2)A retinal vascular segmentation model based on conditional generation countermeasure network is designed.Aiming at the problems of under segmentation,over segmentation and low sensitivity of micro vessels in complex distribution in fundus retinal vascular segmentation.An improved blood vessel segmentation algorithm model based on generative countermeasure network is proposed.The generator part of the model is based on the encoding and decoding structure and the improved U-shaped structure.On this basis,combined with the multi pair encoding and decoding structure in ladder net,the two original improved U-shaped network structures are connected in parallel and expanded into W-shaped network structure,which has been more information flow paths.As a result,the model gets more supplements on the characteristic graph to make use of context information better.The accuracy and efficiency of retinal segmentation are improved.In the discriminator part,the original convolution network added the residual block structure to solve the influence of network degradation.The accuracy of retinal blood vessel segmentation of the model as a whole is improved greatly,through the alternate training with two improved network models.In this paper,the two models are trained and tested on the fundus image DRIVE,STARE and CHASE-DB1 data sets.The results of the experiments reveal that the sensitivities of the two models on the three data sets are 0.8093 and 0.8326,0.8112 and 0.8238,0.8118 and 0.8394 respectively.The suggested two upgraded network models are accurate in segmenting micro vasculature.The sensitivity of retinal blood vessel segmentation pixels has improved greatly,and the network model’s overall effect has improved. |
