Automatic assembly of ocular fundus images

摘要

This paper describes a method of assembling (grouping) ocular fundus images to construct a complete view of the retina. The method uses the structural information contained in the vascular network and is based on finding all possible match locations by comparing blood vessel skeleton segments in pairs of images. The match locations are further analysed at a global level to arrive at the appropriate relative translation and rotation values. To ensure that no discontinuity exists in the complete view, a refinement process is applied to the computed relative translation and rotation values. Typical results obtained from the analysis of various samples yielded maximum translation and rotation errors of 2 pixels and 1° respectively. The former corresponds to a positioning error of 0.6 mm, if the retinal photographs were used for manual assembly. The development of the algorithm and a sample of the results obtained are presented.