Research Of Retinal Blood Vessel Image Segmentation Base On FCN

The fundus retina contains abundant fine blood vessels.The analysis and research of blood vessels in color fundus images is one of the important ways to understand the information of blood vessels in all parts of the body.At the same time,the abnormal changes in the morphological representation of the blood vessels in the retina are often related to some specific eyes and comprehensive systems Diseases are closely related.Segmentation and extraction of retinal blood vessels can provide reliable information basis for clinical diagnosis,prevention and development stage judgment of these diseases,and then have a positive impact on subsequent treatment decisions.In order to improve the efficiency of retinal image processing and avoid the influence of subjective factors and personal experience on the segmentation results,building accurate and rapid retinal vessel segmentation models is of great significance for assisting diagnosis.This thesis aims at the algorithmic characteristics of Fully Convolutional Networks(FCN)in image semantic segmentation,and optimizes and improves from multiple perspectives such as data preprocessing methods,network hyperparameter settings,and model structure.Extensive experiments were performed on the STARE retina dataset.The main work of this thesis includes:(1)The application scenarios and principles of full convolutional neural network in the field of deep learning image segmentation are analyzed in detail.Based on the end-to-end-pixel-to-pixel characteristics of the full convolutional neural network,a reasonable segmentation network structure is constructed and implemented.Based on this,the effectiveness of the full convolutional neural network for the fundus retinal vascular image segmentation is verified.(2)Comprehensive pre-processing methods for data augmentation in common fundus retinal vascular segmentation,use U-Net network to analyze different retinal vascular image data augmentation methods,and compare the effects of different optimization objective functions on retinal vascular segmentation.Five kinds of evaluation indexes are used to evaluate the impact of the data augmentation process and method on the performance and segmentation results of the full convolutional neural network from an objective perspective.(3)Aiming at the problems of slow convergence and low accuracy of the retinal vascular segmentation model,a combination of residual learning,dilated convolution,and batch normalization(BN)optimization is proposed based on the U-Net network structure.Fundus Retinal Vessel Segmentation Method DResU-Net.Residual learning allows the network layer to fit the residual function to facilitate the adjustment of network weight parameters and the transfer of gradients;the expansion convolution expands the receptive field range to extract more abstract feature information while the number of control parameters remains unchanged;Batch normalization is used to regulate the distribution of characteristic data to accelerate convergence.Experimental results show that the improved DResU-Net network has strong generalization ability and accuracy.(4)In order to further improve the segmentation accuracy of the terminal microvessels in the retinal image,a multi-scale feature densely connected down-sampling structure is proposed,which combines context feature information IDU-SSP.The Inception multi-scale convolution feature result is passed as a shared feature of densely connected(Densely Connected Convolutional Networks,DenseNet)in the backbone encoding network to enhance the flow of features in the network.Pyramid pooling is used to extract the context feature information in the data.Layer decoding to obtain better results of blood vessel segmentation.It is verified by comparison experiments that the method performs better on segmentation of whole blood vessels and terminal microvessels.By comprehensively comparing the results of the experiments,the IDU-SSP method proposed in this thesis has a significant improvement in the accuracy of retinal blood vessel segmentation,and can better recognize the local details of the retinal image blood vessels and the overall structural pixels,ensuring network accuracy the case with generalization ability has better practicability.