Research On Retinal Vessel Segmentation Of Fundus Image Based On Fully Convolutional Neural Network

The different forms of retinal blood vessels in the fundus(such as diameter,length,branch,etc.)have very important clinical significance,which can be used to monitor and analyze various cardiovascular and cerebrovascular diseases and ophthalmological diseases.Automatic segmentation of retinal blood vessels in fundus images can effectively help doctors make more comprehensive diagnosis and further treatments for patients with various cardiovascular and cerebrovascular diseases and ophthalmological diseases,and increase the possibility of cure.In the process of research,it was found that it was very difficult to accurately segment the retinal vessels of fundus images because of the influence of different acquisition equipment,shooting Angle and illumination on fundus image acquisition.In addition,the existing models have serious loss of the characteristics and information of the blood vessels during the segmentation process,which results in the poor segmentation effect of small retinal blood vessels.In view of these problems,this article has done the following:(1)The fundus images were processed by means of grayscale processing,normalization,adaptive histogram equalization of limited contrast and gamma nonlinearization to reduce the influence of acquisition equipment,shooting Angle and illumination on the segmentation of retinal vessels in fundus images.And the stepwise overlapping sampling method is used to amplify the data set,which provides sufficient data support for the following image segmentation.(2)Constructed based on improved U-Net segmentation algorithm(RCU-Net).In view of the fact that the characteristics and information of the blood vessels in the segmentation process of the existing models are severely lost,resulting in the poor segmentation effect of small retinal vessels,the residual module and attention mechanism are introduced into the U-Net algorithm.Through experiments on the DRIVE data set,it is proved that after adding the residual module and the attention mechanism,the model can effectively segment the small retinal blood vessels in the fundus image.(3)A segmentation model based on full convolutional neural network and full connected CRF is constructed.In order to make the clinical application effect better,a combination of supervised and unsupervised algorithms are used to complement each other’s advantages.A fully convolutional neural network segmentation algorithm(RCBAM-Net)based on a mixture of residual and convolutional attention mechanism modules is constructed,which reduces the parameters of the segmentation algorithm and improves the efficiency of segmentation.In view of the problems of excessively large receptive fields and insufficient edge constraints in the full convolutional network segmentation results,after the full convolutional network segmentation,the segmentation results are refined by using fully connected conditional random fields.Through experiments,the performance indicators of the model have been improved.