Research On Vascular Segmentation Algorithm Of Fundus Image Based On Improved Level Set

Automatic segmentation of retinal vessels is of great significance to shorten the diagnosis period of diabetic retinopathy.The tissue structure of human eye is complex.In addition,fundus image can be easily mixed with noise due to uneven illumination in the process of image acquisition.These lead to low segmentation accuracy of existing segmentation algorithms for blood vessels.In order to solve this problem,the improved level set retinal blood vessel segmentation algorithm is used to process the fundus image in this dissertation.The processing process is roughly divided into three stages: preprocessing,enhancement and segmentation of blood vessel.1.Aiming at the problems of noise and optic disc in the fundus image,the image is preprocessed.The green channel image is selected as the image to be processed.Gray correction and homomorphic filtering are used to enhance the image contrast and remove the retinal bright noise.In order to avoid the optic disc pixels being wrongly divided into vascular pixels.clustering,connected region analysis,Hough transform,region growing and morphology are used to remove the optic disc.2.Aiming at the problems of the vertigo phenomenon of Retinex enhancement algorithm in the process of dealing with the region with large gray difference,a vascular enhancement algorithm based on wavelet transform and multiscale Retinex is used.The image is decomposed by wavelet transform into low-frequency subband and high-frequency subband.Then the vascular profile information in low-frequency image component and vascular detail information in high-frequency image component is respectively extracted using bowler-hat transform and multiscale Retinex algorithm.After reconstructing vascular profile and details with image entropy as weight,the enhanced vascular image is obtained.3.DRLSE relies heavily on the initial position of the contour and can’t overcome the gray non-uniformity of the image,which result in the insufficient segmentation of the level set model for the micro vessels.The DRLSE is improved by fusing the information of vessel size,phase and linear morphology.The boundary information of the blood vessel located by GAC model is taken as the initial contour.The blood vessel features extracted by two-dimensional Gabor wavelet transform and multiscale multidirectional line detector are used to construct a new edge stop function.At the same time,the CV model is introduced to calculate the local gray energy to build the level set model.Finally,the blood vessels are detected from the fundus image.The accuracy of this model in DRIVE and STARE is 95.15% and 92.68%.Compared with the level set model based on global and local selection,the accuracy of vessel segmentation is improved by 1.93% and 1.77%.