In this project the pattern recognition lab develops algorithms and a software environment for use in multimodal (PET, CT, MR), small animal imaging. Standard algorithms have to be adjusted to the specific challenges in small animal imaging, arising from the small subject size. The work focuses on registration and segmenation techniques.

Image registration techniques allow to correlate different modalities, such that for example marker specificity to the tumor region can be easily evaluated. Both rigid and non-rigid registration algorithms are currently under development. The rigid registration is focusing on speed, accuracy and robustness with respect to transforms that contain only rotations and translations. The nonrigid registration calculates a free form transform, which allows for a direct pixel-by-pixel comparison of the registered images.

Segmentation techniques that are currently investigated are mostly semi-automatic. This allows the user to specify roughly the regions of interest that are then automatically refined by the algorithm. The Random Walk algorithm has been implemented, to facilitate easy and fast segmentation. The algorithm is hardware accelerated on the graphics hardware (GPU) and can yield a segmentation in less than 5 seconds.

Future work will mostly focus on adapting existing algorithms to small animal imaging. In cooperation with the Department of Nuclear Medicine it is planned to identify useful workflows for the evaluation of PET marker performances. These workflows can then be supported by specifically developed software tools.