Design Of Fundus Vessel Segmentation System Based On Improved UNet++ Algorithm

With social development and progress,human living standards have improved,but at the same time,the epidemic of many chronic diseases has become more and more serious.These diseases not only cause direct damage to the body,but their complications can also bring serious consequences such as reduced vision and even blindness.By analyzing the segmented fundus blood vessel images for early inspection and long-term monitoring,the risk of vision loss in patients will be effectively reduced.However,due to the complex structure of fundus blood vessels and low contrast,many existing fundus blood vessel segmentation methods still have problems in segmenting small blood vessels and eliminating noise.The UNet++ algorithm is a deep learning algorithm with good performance.It can segment and extract specific parts of an image,and has a wide range of applications in the field of medical image segmentation.In this paper,the improved UNet++ algorithm is used for fundus image processing,which realizes an automatic system for fundus vessels segmentation and improves the segmentation performance.Firstly,in order to solve the problem of low blood vessel contrast in fundus photos and the dataset contains a small number of pictures,the original data set pictures are first subjected to channel split,CLAHE operation and gamma correction operation to enhance blood vessel contrast,reduce noise interference,and then perform data enhance the operation and increase the number of images in the training set.Secondly,in view of the poor segmentation effect of small blood vessels and the inaccurate positioning of the main blood vessels,firstly,the encoder-decoder networks of different levels are superimposed to form a multi-level network that can extract features of different depths.Then use the nested skip connection to connect the modules of each layer to narrow the gap between different levels of contextual semantic information.And on the feature extraction module,hole convolution is used to change the feature extraction receptive field.Finally,the extracted multi-level feature information channels are connected to obtain multi-level fusion features with different scales of feature information on the feature output layer,so that the deep learning network can learn features with different layers of semantic information.In order to verify the effectiveness of the segmentation system designed in this paper,experiments were carried out on three public standard datasets DRIVE,STARE,and CHASE＿DB1.The experimental results show that the segmentation system based on the improved UNet++ algorithm can accurately extract the details of blood vessels,and the accuracy and sensitivity of the segmentation results reach 98.17% and 89.48% on the DRIVE dataset.Compared with many existing methods for the segmentation of retinal vessels,it has better segmentation effect and can basically meet the requirements of medical auxiliary diagnosis systems.