When registering images of a retinal angiograph, the key feature for the clinician is blood vessel structure. That is also the key to the registration method (up to 36 frames) described and demonstrated in this paper. The major steps in the process are preprocessing, matching, and image warping. This paper discusses the literature for each of these steps (38 references), and the rationale for the methods chosen for this application.

The preprocessing technique is roughly a gradient filter. Motion between frames is determined by hierarchical refinement from global to local, and the warping is applied at each step to achieve successive refinement. Excellent results are obtained in the presented cases, but the authors mention that increased robustness is a goal (a batch of 30 cases resulted in success rates of 90 percent and 30 percent for two types of dye). Another item to address is whether or how image degradation from multiple warping operations can be addressed (techniques do exist to convert multiple warps into a single warp).