The Research On The Processing And Registration Method Of Fundus Images

Fundus images contain a lot of information to reflect the health of the eyes andthe whole body. Eye diseases and various common diseases, such as hypertension,diabetes, cerebrovascular sclerosis and other systemic diseases have different levels ofexpression on the fundus images. So the processing of fundus images related to thesediseases has an important role on the diagnosis and treatment of diseases for doctors.The key technologies include retinal vascular segmentation and rapid automaticfundus images registration.The difficulty for vascular segmentation and registration of fundus imagesmainly from two aspects: First, due to noise, pathology interference and the poorquality of images, it is difficult to extract the retinal vascular and the key featurepoints; Second, it is difficult to register a pair of images taken years apart, which evenacquired with different modalities and scales etc. In this paper, after analyzing thefeatures of fundus images for medical application and do deep research in variousalgorithms on vascular segmentation and registration of fundus images, vascularskeleton is extracted based on improved entropy threshold and Hilditch thinningalgorithm, then we propose a registration algorithm relying on local vasculartree-branch points structure using the vascular skeleton. Specific work is as follows:1. After comprehensive investigating the features of fundus images for medicalapplication, we analyze some commonly used algorithms of vascular segmentationand registration of fundus images. And also we introduce some related basic conceptsfor fundus images processing, especially detail on the algorithms based on SIFT andHarris-PIIFD feature points, experiments and experimental results are given in thepaper.2. We do fundus image preprocessing, including gray processing, noise reduction, region of interest (ROI) extraction, vascular contrast enhancement and vascularskeleton extraction. We use median filter to do noise smoothing and do vascularenhancement based on a combination of morphology and Gaussian filter algorithms.We process the enhanced fundus image using vascular extraction techniques based onimproved entropy threshold, then we obtain vascular skeleton based on Hilditchthinning algorithm.3. We propose a fundus image registration algorithm relying on vascularbifurcation points structure, after removing the pseudo-branch points, we extract thebifurcation points structure which includes a seed point and three connected neighbors.Then we do the first registration relying on the angle and length information of thestructure, and we reuse the improved RANSAC algorithm to remove mismatching,completing fine-registration. Finally we do the wavelet image fusion to complete theregistration. The registration algorithm in this paper uses vascular bifurcation pointsstructures instead of individual feature points, which reducing the number of featurepoints, improving the computational efficiency. At the same time, the characteristicvector of each bifurcation structure consists of the normalized branch angle and length,which is invariant against translation, rotation, scaling. What’s more, using RANSACalgorithm also improves the accuracy of registration. Meanwhile, more natural imagefusion algorithm is used in this paper.