A multi-disciplinary research team at the Medical University of Lübeck is working on an atlas of the human brain within the human neuroscanning project (HNSP). For this purpose, several image modalities have been applied on a prepared male human post mortem brain. The data is particularly derived from histological sections as to obtain specific cellular information. The sectioning process however leads to deformed images which present a major problem for scientists to acquire an insight into the different cerebral structures. This type of distortions can be corrected through elastic matching, a solution that requires a high memory and computational power. Therefore, the team has designed a fast algorithm to be ported on a cluster of 48 Pentium II PCs connected via Myrinet, for parallel implementation.
In fact, the elastic matching method is based on the use of a non-linear partial differential equation (PDE). This approach has the advantage that no supplementary landmarks for the underlying images are needed for the correction, as would be the case in determining coefficients of piecewise linear functions by a least squares condition. Because of the extremely high resolution scans made from the histological brain sections, the PDE can only be calculated with an iterative solver, installed in a parallel manner. Instead of relying on expensive dedicated parallel computers, the Lübeck team is using the "Störtebeker Cluster", which consists of 48 dual 333 Mhz Pentium II nodes interconnected via Myrinet, a Gigabit-per-Second Local Area Network.
The parallel programme runs on LINUX and applies Parallel Virtual Machine (PVM) as the underlying message passing system. PVM enables software to easily create and manipulate tasks on remote computers and communicate efficiently in scalable heterogeneous environments. To correct the deformed images for adequate reconstruction of the data, the parallel implementation has to perform an iterative outer and inner loop. For the latter, the conjugate gradient (CG) method is used. The current system supports only one process per node but the measurement times can be kept reasonable by maintaining the amount of steps in the outer as well as in the inner loop at 50. The team has executed one matching of two 256 x 256 pixels image and another one of two 512 x 512 pixels images.
The more computation nodes are implied, the less calculation time is needed. However, the amount of communication time increases, resulting in a lower rate of efficiency. For small-sized images matching, this may form a problem but in the human neuroscanning project, only large-sized images with a high resolution are used. The parallel elastic matching technique can equally pay good services for other imaging modalities, like computed tomography (CT), positron emission tomography (PET), magnetic resonance imaging (MRI), and electro-encephalograpy (EEG). Thus, in multi-modal image matching of large sized visual data, linear artifacts originating from down-scaling and bilinear interpolation, can be minimized with use of the discrete non-linear elasticity model.
Future work for the Lübeck research team involves the installation of a multi-grid solver. As such, the scientists are heavily interested in the use of a direct solver based on fast Fourier-type techniques (FFT). To this purpose, several modules for solving the system of linear equations are required which have to be matched with appropriate parallelization strategies. For the HNSP project, the team indeed has to reckon with an uncompressed amount of flat bed scanned data of about 700 GBytes and 40 GBytes of episcopic data for a single human brain. The final reconstruction constitutes the basic structure for the integration of functional data based on stochastic mapping as well as for future modelling and simulation studies on a virtual brain. Please, visit the Web site of the Medical University of Lübeck for more details.