Retina Fundus Registration Based On Local Vascular Structure

The retinal fundus image is an important basis for diagnosing of eye disease suchas diabetes, glaucoma, hypertension, coronary heart disease and so on, the diseasewould lead to changes in retinal vascular morphology. The characteristics ofretinopathy usually relate to the symptom of many diseases. Thus it can provideimportant information for doctors to diagnose diseases. The retina image registrationwhich this paper studied is the premise for image fusing and lesion dection, and theregistration result have important auxiliary role for diagnosing many diseases.The difficulty of high-precision registration of retina fundus comes from twoaspects: firstly, some important feature points will be lost when there are present largearea of lesion; secondly, serious nonlinear scale deformation appear in many retinalimages, which shows more severe in the area that far away the center of fundusimages. These factors add the difficulty to calculate the ill-posed registration problem,and cause difficulty to obtain accurate registration results with strong robustness. Lotsof work of retina image registration has proceeded in reported literature, but most ofthem focuse on the normal retina images, and have problem in the registrationefficientcy and accuracy, besides of that, it will cause registration failure in the lesionimage because of missing key feature points when it influenced by noise and/orlesions. According to the characteristics of vascular structures, this paper design aimage registration method based on the local vascular structures and it also adapt tothe lesion images. The main work including in this paper are as follows:1) Based on the Gabor filter, and combing with the boundary inhibition operatorand regional growing algorithm, we solute the problem of extracting vascular skeletonin lesion images.2) The critical factor of registration is obtained vascular matching feature pointswith strong invariance in translation, rotation and scale. Because of the weak power ofanti-scale invariance in isolated feature points, to describe local vascular structures,we identify the adjacent feature points along the vessel direction, and take the relativeposition and distance between them as similarity measure. Due to reflect the mappingrelationship of local vascular structures，these features have strong robustness.3) An algorithm combing with local and global registration technology isproposed against the problem of larger offset appear in part of area caused by nonlinear deformation. First, extract the registration bifurcation structures use thesimilarity measure, then register images with affine transformation or quadraticsurface equation, finally, fine registration is applied to the area in which accuracyregistration is not achieved in original global registration. The algorithm increases theregistration efficiency and accuracy by reducing the matching point set and takingmulti-level registration strategry.Proposed registration method solve the problem that caused by lesions andnonlinear scale deformation in images at some externt. Comparing with other similarmethods, this method present accurate registration for both normal and lesion retinaimages.