Retinal Vessel Segmentation Based On Deep Learning

Retinal blood vessels are the only blood vessels in the human body that can be clearly observed without the use of noxious means.Professional doctors can accurately diagnose multiple diseases by observing the characteristics of retinal blood vessels.However,the retinal vascular structure is complex and often leads to difficulty in observation and risk of missed detection.Therefore,high-precision automated retinal vein segmentation technology is of great significance for clinical medical aid diagnosis.For decades,although a large number of model-based and learning-based methods have been applied to the retinal vessel segmentation problem,in practice,due to factors such as changes in lesions or photographing patterns,these methods have a greater loss of precision in practical applications.It is difficult to meet the expected requirements.In recent years,deep learning has performed well in the field of image processing.Therefore,this paper proposes a multi-scale network algorithm based on deep learning.Its main contributions are:(1)Multi-scale deep supervision network,using a three-layer encoder-decoder structure,using depth supervision to constrain the feature expression of each layer,hopping and merging feature information under adjacent scales,and the boundary optimization module adjusting the network output feature map boundary.(2)Multi-scale semantic fusion network,using a five-layer encoder-decoder structure,based on the deep supervision and jump connection of the multi-scale deep supervision network,using the semantic fusion module to compare each layer of the network The lower layer features are merged,and the semantic information under the high-dimensional features of the network is used to guide the adjustment of the texture information under the low-dimensional features,so that more accurate and consistent blood vessel segmentation results can be obtained.The multi-scale feature optimization module is used to map the characteristics of the network output.For further extraction of vascular information at multiple scales,vascular information at different scales can be extracted simultaneously.In order to verify the effectiveness of the retinal vessel segmentation algorithm proposed in this paper,this paper can achieve high segmentation accuracy by training and testing in several public datasets.