Diplomarbeit Techniques for Accelerating Intensity - based Rigid Image Registration

One important application of computers in medicine is the merging of various sets of patient data coming from different sources. In the case of image registration, the aim is to determine the translation and rotation by which one data set can be spatially aligned with a second one. Volumetric data (e.g. CT data) can be registered to another volume or to a two-dimensional image (e.g. an X-Ray image). An essential part of this process is the ability to measure the similarity between two images. This paper focuses on various techniques for speeding up assessing the similarity of two-dimensional images. In most cases, the registration process requires 2D images to be generated from volumetric data, which can be accomplished much faster by using the capabilities of modern graphics hardware. One acceleration approach therefore consists in also moving the similarity measure computation to the graphics board. The second approach is taking advantage of parallel processing capability (SIMD ”single intruction, multiple data”) of modern processors. The various similarity measures and implementations are benchmarked for a 3D-3D registration process on different data sets. The results show that we can register two volumes very accurately in 20% to 25% of the time needed for a non-accelerated approach.