Covariance-driven retinal image registration initialized from small sets of landmark correspondences

An automatic retinal image registration algorithm would be an important tool for detecting visible changes in the retina caused by the progress of a disease or by the impact of a treatment. Developing such an algorithm is difficult, especially for feature-poor images of diseased eyes. In this paper, a new retinal image registration algorithm is described that bootstraps an estimate of the parameters of a high-order, inter-image transformation model based on just one or two initial retinal image landmark correspondences. Hypothesized sets of initial correspondences are obtained through invariant indexing. For each such set, an initial, low-order transformation covering a small image region is estimated. Sufficiently accurate initial estimates are gradually expanded to a high-order transformation that covers the entire retina using constraints generated by alignment of the vasculature. The expansion and switch in model orders is entirely driven by the covariance matrix of the estimated transformation parameters. The resulting algorithm registers images to accuracies of less than a pixel in just a few seconds.