Ophthalmologic Image Registration based on shape-context: Application to Fundus Autofluorescence (FAF) images

A novel registration algorithm, which was developed in order to facilitate ophthalmologic image processing, is presented in this paper. It has been evaluated on FAF images, which present low Signal/Noise Ratio (SNR) and variations in dynamic grayscale range. These characteristics complicate the registration process and cause a failure to area-based registration techniques[1, 2]. Our method is based on shape-context theory [3]. In the first step, images are enhanced by Gaussian model based histogram modification. Features are extracted in the next step by morphological operators, which are used to detect an approximation of vascular tree from both reference and floating images. Simplified medial axis of vessels is then calculated. From each image, a set of control points called Bifurcation Points (BPs) is extracted from the medial axis through a new fast algorithm. Radial histogram is formed for each BP using the medial axis. The Chi2 distance is measured between two sets of BPs based on radial histogram. Hungarian algorithm is applied to assign the correspondence among BPs from reference and floating images. The algorithmic robustness is evaluated by mutual information criteria between manual registration considered as Ground Truth and automatic one.