Fundus Retinal Vessel Image Segmentation Method Based On Deep Learning

The early detection of blood vessels in the retinal by fundus imaging technology is of great significance for the prevention and screening of related eye diseases.However,it is difficult to achieve rapid detection just by traditional manual detection methods.Therefore,the realization of automatic retinal vessel segmentation is of great value.Limited by the complex structure and changeable morphology of retinal blood vessels,the existing image processing methods cannot well adapt to clinical needs in terms of accuracy and speed.In recent years,medical image segmentation methods based on deep learning have shown great application prospects.In this thesis,based on the detailed study of retinal vessel segmentation methods of domestic and foreign scholars,two deep learning networks based on improved U-Net are proposed for accurate segmentation of retinal blood vessels.The main research work of this thesis includes:(1)We design a multi-scale attention retinal vessel segmentation network based on U-Net(called MA-UNet).MA-UNet is based on the U-Net structure.The encoding-decoding structure and skip connection contained in U-Net can effectively fuse the feature information of each layer and improve the accuracy of segmentation.MA-UNet proposes a multi-scale fusion module and a three-type attention module based on the U-Net structure.The multi-scale fusion module is mainly used to fuse important features at various levels,and retains the spatial dimension information containing high-resolution features,and the three-type attention module can make the network automatically adjust the weight of each channel,thereby strengthening important features and suppressing the expression of irrelevant features.(2)We design a multi-resolution fusion retinal vessel segmentation network based on U-Net(called MR-UNet).MR-UNet proposes a multi-resolution fusion module and a region-sensitive attention module based on the U-Net structure.Among them,the multi-resolution fusion module includes the encoder fusion sub-module and the decoder fusion sub-module,which are located at the end of the encoder and decoder structure,respectively,and are mainly used to fuse the characteristic information of different resolutions of each layer of encoder and decoder.The regionally sensitive attention module consists of a channel attention submodule and a spatial attention submodule,which is used to train the network to improve the network’s attention to the vessel region and ignore irrelevant regions.In this thesis,two publicly available retinal vessel datasets,DRIVE and CHASE＿DB,are used to verify the segmentation performance of the above two networks.The experimental results show that the two network models designed have a certain segmentation ability for retinal blood vessels.At the same time,the ablation experiment also proved the effectiveness of the individual functional modules.