Research Of Retinal Image Registration Methods

Retinal image registration is a very important field in image analysis. Because of the special structure of the retinal, one captured image cannot meet the requirement of clinical applications, image registration and image stitching are significant both for diagnosis and treatment. By registering retinal images captured in different time, doctors can diagnose fundus diseases better, such as diabetes, glaucoma, macular degeneration, and so on. Latest researches show that these diseases can present morphological changes of fundus retinal vascular. This paper focuses on the research of three retinal images registration algorithms and this study has clinical significance for early diagnosis of eye diseases.Firstly, it's the Maximum Mutual Information image registration algorithm. Mutual Information for image registration has achieved great success, and till now most researchers have accepted it as a good registration guideline, especially for multi-mode image registration The advantage of this approach is that without any image pre-processing and segmentation, a good registration result can be got. However, the disadvantage is that it is not suitable for image registrations with large changes in rotation and scale.Secondly, this paper mainly focus on retinal image registration methods based on feature points. Two methods for feature point extraction in this paper, one is SIFT (Scale Invariant Feature Transform) algorithm, and the other is eight -- neighborhood method. The former method can extract all the feature points in images, while the latter can extract the bifurcation points of vascular centerlines as feature points. ICP (Iterative Closest Point) algorithm is used for image registration based on the feature points.Finally, a new image registration algorithm based on retinal structural feature is proposed. The structural feature consists of one main bifurcation point and its three neighbor bifurcation points. By finding optimal matching characteristic structures, precise image registration is achieved. However, the good result of this algorithm depends on good extraction of vascular centerlines. This algorithm works easier than the previous two methods, and is easy to be implemented. Experiments show that the method is effective.