Research On Retinal Vascular Image Segmentation Based On Deep Learning

The eye is the most commonly used organ to capture information in daily life,and most of the knowledge or memory we have is from the eyes to perceive the external information,stored in the brain after cerebral cortex feedback,so its importance is selfevident.With the development and progress of science and technology,eye health problems caused by improper personal eye hygiene have become a common problem urgently to be solved in the medical community.Effective prevention or treatment of eye diseases should be based on accurate diagnosis.Retinal blood vessels of the fundus are the only tissue structure in the human body that can be directly observed in a nondestructive way,accurate segmentation of the retinal blood vessels in clinical treatment can make the medical staff more intuitive by observing the changes in the morphological structure of retinal blood vessels of the fundus,so as to prevent or further diagnose related diseases.Therefore,based on the high accuracy of retinal blood vessels segmentation method research,has extremely important clinical significance and practical value.Traditional way of manual segmentation depends on and is limited by the doctor richness of the theory of knowledge,lead to split work is difficult to guarantee accuracy and low efficiency.The early automatic segmentation methods mostly rely on the image processing technology of shallow feature extraction,and can not obtain the deep features with high ductility and robustness.Deep learning can realize the characteristics of end-to-end,high-precision automatic segmentation,and its application in the task of retinal vascular image segmentation can significantly improve the segmentation accuracy.Based on the characteristics of retinal blood vessel image,this paper adopts deep learning method to carry out the research task of retinal blood vessel segmentation.The main research contents of this paper are as follows:1.Research the two traditional deep learning models,U-Net and FPN,they are respectively applied to the retinal blood vessels segmentation tasks.The real data set of fundus image--DRIVE was used to verify the influence of data set clipping method,enhancement strategy and other processing work on model segmentation accuracy through experimental comparison,and the data preprocessing method more suitable for deep learning model was screened out.At the same time analyzed the performance of two kinds of basic model.2.Study the retinal vascular segmentation method based on the multi-scale fusion feature pyramid.In the context of the original structure of FPN resolving power shortage phenomenon,through the adoption of different rate of dilation convolution,build the characteristics of multi-scale fusion model of the pyramid fusion algorithm.The improvement of the segmentation performance of the algorithm model is verified by experiments,and the focal loss function is introduced in the model experiment to suppress the model migration caused by the imbalance of positive and negative samples caused by random clipping.3.Study the deeper U-Net segmentation method based on the context fusion module.For U-Net structure not robust enough and details due to the lack of semantic information,this paper constructs the deeper U-Net structure,and combines with the context fusion module to extract stronger feature information.The experimental results show that this structure can control the details more effectively.