REGISTRATION AND 3-D RECONSTRUCTION OF A RETINAL FUNDUS FROM FLUORESCEIN IMAGES SEQUENCE
Tae Eun Choe
Registration and reconstruction of a retinal fundus from fluorescein images are important to diagnose retinal diseases. This is difficult due to intensity changes in fluorescein images and the near-planar retinal surface, which has narrow depth variations. This study proposes methods of the registration and the 3-D reconstruction of a retina from a sequence of images. We first present a method locating vesselsí» bifurcations, so called Y-features, which is a robust geometric feature in the retinal image. Subsequently, fitted Y-features are matched across images using mutual information to register their unordered colors and fluorescein images globally. Second, using matched Y-features, the epipolar geometry for a near-planar surface is robustly estimated by a plane+parallax approach and dense disparity map is estimated using mutual information criterion. Finally, the 3-D structure of the near-planar retinal fundus is reconstructed in Euclidean space by introducing the three-pass bundle adjustment method. Our experimental results validate the proposed method on a set of difficult fluorescein image pairs and other near-planar images.