Research On Retinal Vascular Segmentation Technology Based On Deep Learning

In clinical practice,retinal images can reflect the symptoms of various diseases.Extracting retinal blood vessels from retinal images can provide a basis for ophthalmologists to screen and diagnose related diseases.Previous artificial retinal vascular segmentation methods have high costs,low efficiency,and strong subjectivity,which cannot meet clinical needs.Therefore,it is particularly necessary to design an accurate,objective,and efficient automated retinal vascular segmentation method.However,the formation of retinal images is easily influenced by various environments,making it difficult to distinguish blood vessels from the background;In retinal images,the morphology and structure of blood vessels are complex,with a large number of small blood vessels distributed,and some areas still have lesion interference.These factors make automatic extraction of retinal blood vessels very challenging.This article proposes two deep learning based retinal blood vessel segmentation methods by conducting in-depth research on the basic methods of existing blood vessel segmentation.(1)IAU-Net segmentation method based on attention and multi-scale feature aggregationThis thesis proposes an IAU-Net segmentation method based on attention and multi-scale feature aggregation to address the issue of high interference in retinal images and complex and variable vascular structures.Firstly,by performing a series of preprocessing on the fundus dataset images,the readability of the fundus images is improved and the amplification of the fundus data is achieved;Then design the attention and multi-scale feature aggregation network IAU-Net.First,the multi-scale residual convolution module is used to reduce the loss of information and effectively expand the Receptive field,so that the network can aggregate multi-scale vascular feature information and extract more feature information of small blood vessels.Second,the Drop Block layer is added to the network model,and the Drop Block is used to close some neurons in the network to prevent the model from overfitting.The third is to use a joint attention module in skip connections,which effectively suppresses the multi-level irrelevant region features of network integration while combining shallow and deep features of the network,highlighting the vascular features required for each layer of the network,and promoting more accurate recognition and segmentation of blood vessels;Finally,experiments were conducted on two publicly available datasets,DRIVE and STARE,to verify the excellent performance of the method in vascular segmentation.(2)SRTrans CNN Segmentation method based on dual branch network.This thesis proposes a segmentation method based on SRTrans CNN dual branch network to address the issues of mis-segmentation,unstable segmentation results,and weak generalization performance of existing segmentation methods in vascular segmentation tasks.Firstly,perform targeted preprocessing on the image;Subsequently,a dual branch network model of Transformer and Convolutional Neural Network(SRTrans CNN)is constructed.This model designs a convolutional branch network with fine-tuning U-Net as the network backbone to enhance the Semantic information of local blood vessels.Simultaneously designing another Transformer network branch and proposing an improved SR Transformer network for medical fundus images,utilizing the Transformer’s self-attention mechanism for remote modeling to enhance global contextual connections in vascular images.SRTrans-CNN then uses a feature aggregation module to aggregate the feature information of the two branches,enriching the global and local feature information of blood vessels extracted by the network;In addition,a hybrid segmentation loss function is introduced for the two-branch network model to enhance the sensitivity to vessel categories.The experimental results show that this method can effectively distinguish foreground and background information both globally and locally,alleviate mis-segmentation problems,and effectively segment blood vessels with lesion information.The method has strong robustness and generalization performance.