Research On Segmentation And Registration Of Blood Vessels In Multi-modal Retinal Images

Diabetic retinopathy(DR),age-related macular degeneration(AMD)and glaucoma are the leading causes of visual impairment and blindness of aging population.Retinal fundus image is an important basis to help diagnose retinal diseases,so the study of retinal fundus images is very important.Retina is the only part of the human body that allows direct non-invasive visualization of its anatomical components.Automatic segmentation of blood vessels in fundus images can help doctors to analyze and diagnose diseases,which is one of the most important problems to be solved in medical image processing.However,due to the complicated structure of blood vessel network,the low contrast between blood vessels and background,the uneven distribution of fundus images and the influence of image noise,the segmentation of retinal blood vessels faces a great challenge.At the same time,the registration of multimodal retinal fundus images can provide comprehensive information for ophthalmologists,which is good for the diagnosis and treatment of retinal diseases.The difficulties of retinal image registration are:(1)The images to be registered have lesions,which makes it difficult to extract feature points.(2)There is a serious nonlinear deformation in fundus imaging.Due to these two reasons,the conventional method is difficult to achieve high-precision registration,and the robustness is weak.On the other hand,the current research on multimodal retinal image registration is relatively few.Based on the above problems,multimodal retinal image registration is studied based on the basic features of blood vessels.The main research contents of this paper include the following two aspects:(1)Aiming at the problem of incomplete segmentation of blood vessels and poor continuity in existing retinal vessel segmentation methods,a method of vessel segmentation based on random forest is proposed in this paper.Firstly,a 25-dimensional feature vector is constructed for each pixel in retinal fundus image,including Gabor,grayscale and gradient,invariant moment,gray level co-occurrence matrix,phase consistency,convolution kernel function and Hessian matrix.Then the random forest classifier is used to classify the blood vessel pixels and the background pixels to get the initial blood vessel segmentation results.Finally,the small,severed blood vessels are repaired based on the high and low probability of random forest and connected domain,thereby improving the accuracy of vascular segmentation and enhance vascular continuity.(2)Aiming at the problem that the existing retinal image registration algorithms are not ideal due to the deformation of the fundus image or the low image quality,we proposed an algorithm for registration of OCT fundus images with color fundus photographs based on invariant features.Firstly,blood vessels are extracted based on the vessel ridge branch,and the feature matching points are extracted according to the overlapping coefficient of vessels.Then the quadric surface model is used to estimate the transform matrix coefficients to achieve registration.Experimental results show that the proposed algorithm has strong robustness and high registration accuracy for images with noise,lower contrast,light changes and age-related macular degeneration.