Research On The Algorithm Of Blood Vessel Image Segmentation Based On Deep Learning

Blood vessel is one of the most important organs of human body,and its condition is closely related to many potential human diseases.Providing help for the diagnosis of many diseases of the blood vessel image,the segmentation is an important step in medical imaging analysis,research and diagnosis.The segmentation algorithm of blood vessel image based on deep learning and convolutional neural network(CNN)has excellent image feature extraction performance and achieved important breakthroughs in the task of blood vessel image segmentation.The segmentation accuracy and performance far exceed the traditional algorithms.Based on deep CNN and the technology of digital image processing(DIP),the segmentation algorithms of blood vessel image are studied in this paper.Firstly,the characteristics of blood vessel images are studied,and for enhancing blood vessel images,a set of preprocessing algorithms are proposed.Secondly,two kinds of segmentation units are designed,convolution segmentation unit and deconvolution segmentation unit: convolution unit uses multi-layer convolution to extract different level features of blood vessel and background,while deconvolution unit fully combines the feature extraction ability of convolution and the feature decomposition and learning ability of deconvolution.Thirdly,based on the segmentation units,the CNet and CDNet models are designed in this paper.The CNet uses the convolutions to extract and learn the expression of vascular and background features.The CDNet combines the convolution and deconvolution to extract,decompose and learn the image features.Both models adopt the U-shaped structure and residual design,which ensure the combination of global information of high-level features and local information of low-level features,and improves the segmentation accuracy of blood vessel images.Finally,the experiments based on two kinds of segmentation mode,image blocks and end-to-end,are carried out for the two kinds of segmentation models.The experimental results prove that: compared with existing algorithms,the segmentation performances of the two kinds of models proposed in this paper are more precise.On the datasets of DRIVE,STARE and CHASEDB,the AUC(≥0.983 on DRIVE,≥0.9939 on STARE,≥0.9811 on CHASEDB),accuracy(≥0.962 on DRIVE,≥0.9721 on STARE,≥0.9558 on CHASEDB),sensitivity(≥0.7675 on DRIVE,≥0.8765 on STARE,≥0.7484 on CHASEDB)and specificity(≥0.9852 on DRIVE,≥0.9772 on STARE,≥0.9647 on CHASEDB)of the both two kinds of models surpass the existing segmentation algorithms,and outperform state-of-the-art techniques.Among them,the end-to-end algorithms not only surpass other algorithms in performance,but also greatly surpass the patch-to-patch algorithms in running time.In this paper,a variety of different blood vessel segmentation CNN models are studied and designed from different directions.The segmentation performances of these CNN models are quite accurate with fast running speed,high efficiency and convenient application.The study of this paper has great important research value and application value to the field of blood vessel image segmentation.