Image registration acceleration on the Cell B.E.

Project description

Rendering  with emphasized blood vessels
Rendering with emphasized blood
vessels. The blood vessels were
identified using atlas-based segmen-
tation combined with region growing.
Registration of two 3D medical images is a key computational tool for many critical clinical applications. It brings two images of an object into alignment [4]. This allows, for instance, fusing two brain MRI scans that were acquired at different times during a treatment in order to inspect treatment success. There are a number of possible reasons for misalignment: The patient may be positioned differently during the acquisition of the two images or the images may have been acquired with different devices or settings. Linear or rigid registration attempts to compensate deformations such as translations and rotations. However, more complex models are needed to compensate misalignment caused, for example, by soft-tissue deformations or respiratory motion. These nonrigid transformation models may have thousands of parameters, and processing typical 3D images can take several hours. We are working on accelerating nonrigid image registration on the Cell/B.E. processor in order to enable the usability for time critical applications [5].