Research On Fundus Retinal Vessel Segmentation Algorithm Based On Deep Learning

In clinical ophthalmology,fundus retinal images are extremely important for ophthalmologists to diagnose and treat fundus retinal diseases.As one of the most important components of the fundus retina,blood vessels can reflect the progression and condition of many ocular and systemic diseases.However,due to the complex and diverse shape of retinal vessels,numerous bifurcations,low contrast with the fundus background,and the time-consuming and laborious process of manual segmentation,existing fundus retina datasets often have a low volume of data,making retinal vessel segmentation tasks highly challenging.Therefore,this thesis proposes the following two methods:1.PAU-Net,a retinal vessel segmentation network based on attention mechanism and multi-scale feature fusion.In PAU-Net,serial skip connections and parallel multi-branch fusion are used to improve the accuracy of the network in vessel segmentation.A pyramid pooling module is introduced to replace max-pooling downsampling in the U-shaped encoder-decoder network to enhance feature extraction of the network in the region of interest.Furthermore,an attention gate is used to concatenate and extract features for skip connections and upsampling to further enhance the network’s ability to extract features in the region of interest.2.DAU-Net,a retinal vessel segmentation network based on dense connection and self-attention convolution.In DAU-Net,a dense connection is applied between all unconnected layers of the U-shaped encoder-decoder network to form a dense connection network.The attention convolution module ACmix is then used to replace the original convolution module in the network to extract features in the region of interest and suppress features outside the region of interest,thus improving the network’s robustness in vessel segmentation.Experimental results show that PAU-Net and DAU-Net can effectively segment retinal vessels in publicly available datasets,such as CHASE＿DB1,DRIVE,and STARE,with better performance than other segmentation algorithms in recent years.