Research On Vessel Segmentation In Fundus Image Based On Convolutional Neural Network And Conditional Random Field

Related statistics show that the number of blind people in the world is increasing year by year.Among them,eye diseases and some other systemic diseases such as diabetes,hypertension and arteriosclerosis are the main causes of blindness.Screening and diagnosis early in the disease will greatly reduce the likelihood of blindness.Fundus vascular morphological structure is stable,but the lesion will lead to changes in retinal diameter,angle,curvature and branch structure.Doctors can assist in the diagnosis of related diseases by analyzing changes in the structure of fundus retinal vessels,and accurate retinal vessel segmentation is a prerequisite and guarantee for effective disease diagnosis.So far,although many experts and scholars have proposed a large number of efficient algorithms for fundus segmentation of the fundus,the fundus structure of the fundus is complex and often affected by low contrast and lesion area.The segmentation of fundus retinal vessels is still challenging.In order to effectively segment the retinal vessels,a fundus image segmentation method based on convolutional neural network(CNN)and conditional random field(CRF)is proposed in this paper.Firstly,we use a multi-scale convolutional neural network to obtain a probability map to achieve a rough segmentation of retinal vessels.Due to the complex structure of the retinal vessels,consideration of details and high-level information is required.This article uses a multi-scale network structure to fuse rich multi-scale and multi-level information.By this way,we can pay attention to retinal vessel details while identifying the structure of the retinal vessels.At the same time,we find that when the network converged,there were still some retinal vessels that are not identified.In order to emphasize the segmentation of these hard samples,we propose an improved cross-entropy loss function.By ignoring the loss of the easy-to-segment samples,the network is more focused on the learning of hard samples.Secondly,in order to smooth the rough-segmented retinal vessel probability map,we use a fully-connected conditional random field to refine the probability map.The fully-connected conditional random field takes the segmentation task as a problem which needs to obtain the marker value through an observation value.The segmentation result is determined by solving the maximum posterior probability.It makes full use of spatial context information.And that not only smooth the edges of retinal vessels but also overcome the effects of noise to some extent.We define the retinal vessel probability map obtained by the convolutional neural network as unary potential function of the conditional random field model and obtain the final refined binary segmentation result.Finally,we conducted a series of experiments on two common datasets,DRIVE and STARE,to evaluate the segmentation performance of the algorithm.Among them,comparative experiments on multi-scale network structures show that after using multi-scale information,the segmentation performance for small retinal vessels is improved,and the segmentation sensitivity is improved.The contrast experiment on loss function shows that after improving the loss function,the algorithm is further improved and the segmentation effect of a part of edges and small retinal vessels is improved.The experimental results of the overall performance of the algorithm in the two common datasets further show that our method outperforms most mainstream methods in terms of sensitivity and accuracy.