Research On Fundus Image Registration Method Based On PixelNet And K-core

Retinal diseases are manifested on the retina as maculopathy,bleeding spots,exudates and other lesions.These retinal diseases can be diagnosed by observing fundus images.The retinal area is large.One fundus imaging cannot completely obtain all the areas.It is often necessary to capture multiple fundus images to meet the clinical diagnosis requirements.By registering fundus images taken at different angles,the doctor can more accurately check and diagnose various retinal diseases.Therefore,the research of fundus image registration has important clinical significance for early auxiliary diagnosis and treatment of retinal diseases.This thesis presents the fundus image blood vessel segmentation method based on the improved Pixel Net network firstly.Based on the Inception-v1 module,three enhanced convolution modules are proposed and applied to Pixel Net networks.Combining pixel hierarchical sampling and enhanced convolutional structure can better extract pixel features and obtain better fundus segmentation results.Based on the segmented blood vessel,this thesis proposes a fundus image feature point registration method based on k-core decomposition.A directed graph(Di Graph)is established based on the bifurcation points and endpoints of the blood vessel.The k-core at k = 2 is used to decompose the overall structure of the fundus blood vessels to obtain relatively stable registration feature points.And then establish SIFT feature descriptors for registration using the extreme value points detected in the Gaussian difference space.Finally,the image fusion is achieved based on the registration results.To test 15 pairs of fundus images of FIRE fundus image registration database by the proposed registration method.And the results were compared with results of other representative method.The test result show the variance of the registration accuracy of the 15 fundus image pairs is smaller using proposed method.And the average registration accuracy is 91.96%,which is higher than other representative methods.Through the details comparison of the fusion result images,it can be seen that the vascular structure of the fundus image can be better fused after registered using proposed method.The experimental results prove that the fundus image registration method based on Pixel Net and k-core has good robustness and high registration accuracy rate.The proposed method can achieve more accurate automatic registration of fundus images.