Many tasks in medical image analysis require the fusion of two images, for example to combine information provided by different image acquisition devices or monitor disease progression over time. Image registration is the process of aligning two images, such that corresponding points can be related. For this purpose, one image is deformed to match the other one. Rigid registration techniques try to achieve an alignment by scaling, rotating and translating one of the images. Rigid registration is only adequate for special cases because usually the anatomical structures in the images are not rigid and therefore more complex nonrigid transformation models are required. A problem with nonrigid registration methods is their high computational cost yielding registration times in the order of hours for typical 3D images. While this is inconvenient for some applications, it is prohibitive for others, especially in the context of intraoperative scenarios.

This dissertation presents a nonrigid registration algorithm and implementation achieving sub-minute runtimes on a system based on two Cell/B.E. processors. This processor contains multiple processor cores on one chip and is designed for computationally intensive workloads. A speedup of more than 2500x was achieved compared to a sequential open-source implementation running on a general-purpose processor. Although optimized for the Cell/B.E. architecture, the presented algorithm also achieves a high efficiency on more recent general-purpose multicore architectures. A scalability analysis shows that the algorithm has the potential to exploit future architectures with more cores.

The algorithm bases on the B-spline transformation model, which has been applied successfully to a wide range of nonrigid registration problems. Furthermore, it uses mutual information, which is probably the most common similarity metric for multimodal image registration. Mutual information allows registration of images which were obtained from different acquisition devices.