Evaluation of an Automatic System for Simulator/Portal Image Matching

User guided systems for X-ray/portal image matching are commonly employed in radiation therapy for position verification and correction. To enable extended clinical support a fully acceptable solution must be generic, accurate, fast and automatic. In this paper a new registration system using the structure tensor is compared with the performance of experienced radiographers. It is shown that local extremum lines (emanating from bone ridges) are invariant with respect to the photon energy and that they can be detected without a priori knowledge. Registration based on such creases is validated with a large database of images from anterior irradiation fields for prostate treatments. It appears that the precision of the structure tensor approach is significantly better than human observers. The execution time is less than 10 seconds while human supervision can be restricted to detect outliers only.