Design Of U-Net Segmentation Optimization Model For Retinal Blood Vessel Image

The characteristics of human retinal vessels are closely related to a variety of diseases including ocular pathologies and often used as an important tool for diagnosing cardiovascular diseases such as diabetes,hypertension,and coronary heart disease,which have great research value for clinical medicine.Traditional segmentation of blood vessels in fundus retinal images relies on manual segmentation by doctors,which is not only time-consuming but also influenced by personal factors and cannot meet the requirements of large-scale blood vessel image segmentation.With the development of image processing technology,the use of computers to implement efficient,stable and accurate fundus retinal image segmentation algorithms plays an important role in improving the efficiency of physicians in the diagnosis of related diseases.Researchers have currently proposed numerous fundus retinal image segmentation algorithms,but the segmentation algorithms still face challenges and have much room for improvement due to the varying thickness of fundus blood vessels,complex patterns and interference of lesion areas.Based on this,this paper designs two fundus retinal image segmentation algorithms based on U-Net network with improvement and experimental validation on the public datasets DRIVE,STARE and CHASE＿DB1,and achieves good segmentation results.The main works are as follows.Aiming at the problems of bright and dark variable fundus retinal images,low contrast between background and blood vessels and noise interference,a retinal blood vessel image segmentation network based on cross-layer weight feature fusion is designed in this paper.The network firstly incorporates a cross-layer spatial weight feature extraction module,aiming to share the semantic information-rich spatial weight features from the deeper network to the shallow layer during the hopping connection of U-Net and enrich the semantic information of the shallow layer features.Second,the convolutional blocks in U-Net are replaced with bottleneck residual blocks to reduce gradient disappearance and model computation.Finally,a hybrid loss function consisting of a cross-entropy loss function and a Dice loss function is introduced to address the disproportion of vessels to background that causes the network to tend to recognize the background.The experimental results show that the network outperforms the advanced methods of contrast in terms of accuracy and sensitivity,with Accuracy(Accuracy)of 97.37% and Sensitivity(Sensitivity)of 84.89%in CHASE＿DB1,which is more advantageous for dealing with regions with low contrast.For the features of complex retinal vascular patterns and large span of planes,a segmentation network based on multi-scale feature fusion is designed,which adds a context enhancement module for each layer after the decoder based on cross-layer weight feature fusion,which mainly consists of multiplexed dilated convolution to obtain multi-scale contextual information in each layer of the decoding stage.Secondly,the retinal images of different scales are fed into each layer encoder with matching sizes in the encoding stage to enrich the multiscale information in the encoding stage.Finally,multi-scale supervision is introduced for the context enhancement module in each layer of the decoding stage so that the loss function supervises to all decoder outputs of different scales.The accuracy(Accuracy)and sensitivity(Sensitivity)of this network reach 97.76% and 82.99%,respectively,on the DRIVE dataset,achieving better segmentation results at fine vessel edges compared with the compared state-ofthe-art network.