Research On Retinal Vessel Segmentation Methods Based On U-Net And UNet++

With the aging of the population and the change of lifestyle,the incidence of eye diseases is increasing year by year.The detection and analysis of retinal blood vessels play an important role in the screening and diagnosis of many ophthalmic diseases.Fundus retinal images can provide rich information about pathological changes.Various pathological changes directly reflect a variety of eye diseases,and also reflect the health status of other organs to a certain extent.According to the width,curvature,branch shape and angle of blood vessels,it can effectively help doctors diagnose diseases such as ophthalmology and internal medicine.The artificial segmentation method is time-consuming and laborious.At the same time,the topological structure of blood vessels in fundus images is complex and changeable,and the shapes are different.It is difficult to accurately analyze and judge retinal blood vessels by artificial segmentation method alone.With the help of automatic segmentation technology,the accurate segmentation of retinal blood vessels can be realized,the labor burden can be reduced,and the clinical diagnosis efficiency of ophthalmic diseases can be improved.Therefore,this study uses the existing deep learning framework and model to achieve accurate segmentation of retinal blood vessels for retinal blood vessel fundus images.The main contents of this study include :(1)In order to solve the problem of insufficient segmentation of retinal small blood vessels and low accuracy and sensitivity,an improved model based on U-Net architecture is proposed.This model proposes an improved hollow space pyramid pooling method,which can increase the receptive field of the network through multi-scale feature extraction,thus helping the model to identify various forms of blood vessels.Deformable convolution is also used to capture the local structure information in vascular images.Through the nonlinear sampling ability of deformable convolution,the local structure information in vascular images can be better captured,and the accuracy and sensitivity of vascular segmentation can be further improved.In addition,in order to improve the generalization ability of the model,Drop Block technology is used to avoid overfitting.In this paper,the improved algorithm is verified on the public fundus dataset.The accuracy and sensitivity on the DRIVE dataset are 95.73 % and 80.79 %,respectively.The accuracy and sensitivity on the STARE dataset are 96.92 % and 81.87 %,respectively.(2)In order to segment retinal blood vessels more accurately and improve the ability of the algorithm to identify microvessels,an improved UNet + + model is proposed.The model designs a residual spatial attention convolution block to improve the vascular weight and make it better segmented in the spatial dimension.Through the spatial attention mechanism,the network ’s focus on the image can be more focused on the vascular area,and the residual structure is used to solve the problem of gradient disappearance caused by the deepening of the network depth.In order to further improve the segmentation accuracy of retinal blood vessels,the Soft Pool pooling method based on Softmax weighting is used to retain the detailed information in the image.This pooling method can better capture the features in the image and improve the accuracy of the segmentation results.In addition,in the model training,a loss function is designed to reduce the over-fitting phenomenon,and the number difference between positive and negative samples is balanced to avoid the over-fitting problem.In this way,the model can be trained more effectively.The accuracy and sensitivity of the improved model based on UNet + + are 95.76 % and 81.20 % on the DRIVE dataset,97.05 % and 83.18 % on the STARE dataset,and 96.55 % and 81.83 % on the CHASE DB1 dataset.The experimental results show that the two segmentation algorithms can segment more detailed and complete vascular structures,and achieve better performance than most existing algorithms.