Research On Methods For Retinal Blood Vessel Detection

Automatic vessel segmentation in fundus images plays an important role in screening,diagnosing,treating,and evaluating various cardiovascular and ophthalmic diseases.However,retinal vessel segmentation has become a long-standing challenge due to limited well-annotated data,variable vessel sizes,and complex vessel structures.With the development of retinal blood vessel segmentation methods,higher requirements are put forward on the algorithm for the problems of poor blood vessel connectivity,loss of small blood vessels and additional prediction in retinal blood vessel segmentation,as well as new blood vessels,exudates,and hemorrhages in retinopathy.By analyzing the above problems,two retinal blood vessel segmentation algorithms are designed in this paper.（1）Retinal vessel segmentation algorithm based on self-attention fusion.The network consists of a lightweight encoders-decoders and a self-attention fusion module.Encoders-decoders are used to extract features and classify them.The self-attention fusion module is introduced between the model encoder and decoder from the spatial and channel dimensions.The two self-attention modules of the channel and spatial annotation are used in parallel to generate expressive features of the attention perception.The interclass discrimination and intra-class responsiveness are enhanced to learn the rich structural hierarchy of retinal vessels,enabling the model to accurately classify vessel structures from the background and foreground.（2）Retinal vessel segmentation algorithm based on cascaded mini U-Net.Compared with U-Net,mini U-Net is a lightweight network.The cascaded mini U-Net is composed of two mini U-Net iterations.The first mini U-Net performs rough segmentation of retinal images,and the second mini U-Net is a refinement unit to solve the incomplete segmentation.Since there is no significant difference between the blood vessels and the background in the retinal images,a spatial attention module is introduced into the model,causing the task target to be focused on the blood vessel pixels.Areas that contribute more to the results are paid attention to,improving the utilization of spatial dimension features.In addition,the Weight-Share Drop Block convolution block is used in the network to replace the original convolution block in U-Net to avoid network overfitting and accelerate network convergence.The blood vessels are completely segmented by this network and have strong connectivity,which can predict the new blood vessels well..The models given in this paper are trained and tested on the public datasets DRIVE,CHASE＿DB1 and STARE and achieved excellent results,verifying the effectiveness of the model architecture and related modules.In addition,compared with the retinal blood vessel segmentation algorithms in recent years,the two models presented in this paper have achieved excellent results in evaluation indicators such as accuracy,sensitivity,and specificity.At the same time,the cascaded mini U-Net is only tested on the Messidor and ORIGA-650 datasets to verify the generalization ability of the model on data with different distributions.From the visualization results,the cascaded mini U-Net has strong generalization ability.