Elastic registration of brain images on large PC-Clusters

The aim of the human neuroscanning project (HNSP) is to build an atlas of a human brain at cellular level. The database was obtained by a variety of image modalities and in particular histological sections of a prepared brain. As the preparation leads to linear and non-linear deformations of the tissue, reconstructing the essential information out of deformed images is a key problem within the HNSP. Our approach of correcting these deformations is based on an elastic matching of the images. Therefore, a parallel implementation was used, since the problem in general is computational expensive and for very large scale digital images a huge amount of data has to be processed. As these requirements are in the range of today's grand challenges, a large PC-Cluster was used to provide the performance demands. The measurements and results presented here were obtained on a cluster of 48 Dual SMP platforms connected via a Myrinet network.