A registration algorithm of eye fundus images using a Bayesian Hough transform

Image registration is the real challenge of retinal image analysis. Temporal registration is necessary in order to follow the various steps of a disease and to measure the evolution of some lesions. This is particularly true for patients developing diabetes, the first cause of legal blindness in most occidental countries, where the number and the turnover rate of lesions such as micro-aneurysms is related to the gravity of the illness. This paper presents an algorithm for temporal registration of retinal images based on point correspondence. The algorithm has been applied to the registration of fluorescein images (obtained after a fluorescein dye injection). The vascular tree is first detected in each image and bifurcation points are labelled with surrounding vessel orientations. An angle-based invariant is then computed in order to give a probability for two points to match. Then a Bayesian Hough transform is used to sort the possible matchings with their respective likelihood. A precise affine estimate and a score are computed for most likely transformations and the best transformation is chosen.