An advantage of the High-Voltage Electron Microscope (HVEM) is that it can obtain images from thicker specimens that contain substantial three-dimensional structure. Though HVEM images represent a projections through the specimen, information from images taken in different orientations may be tomographically combined to obtain a three-dimensional description. in the single axis tilt procedure, a series of digitized images are acquired as the specimen is tilted in small angular increments about a fixed axis. The tilt series is then processed to derive a digital volume dataset which may then be subjected to image processing, rendering and visualization techniques to reveal the 3D structure of the specimen. Tomographic algorithms require extensive computation and considerable processing time on conventional workstations in order to reconstruct HVEM volumes (typically 512 x 512 x 128 voxels). Significant gains in processing time are anticipated by implementing tomography algorithms on massively parallel machines. In the past year we have made progress in implementing the commonly used single axis tilt, Rweighted backprojection (RWB) algorithm and two iterative reconstruction methods, algebraic reconstruction (ART) and simultaneous iterative reconstruction (SIRT). Iterative techniques promise improved reconstructions by facilitating the incorporation of information that con strains the nature of the specimen under reconstruction. Generally, these reconstruction methods iteratively correct the backprojected volume using an error measure derived from a comparison of the projections from the reconstructed volume and the corresponding tilt data images. These algorithms are often performed after an initial RWB pass, and many iterations may be required to achieve optimum results. Both the RWB and iterative procedures are relatively easy to parallelize because each plane of the volume orthogonal to the tilt axis may be processed separately by an identical algorithm. We have re-engineered conventional workstation software implementing the RWB, ART and SIRT algorithms, and implemented initial versions of these tomographic procedures on the Intel Paragon. Successful reconstructions of volumes have been produced using up to 128 nodes on the Intel Paragon. Verification runs have established that tomographic reconstructions produced by our parallel code are concordant with those obtained using conventional (serial) algorithms. The present implementation results in a modest speedup in comparison to the serial implementation. For example, a reconstructionof a 5l2x64x312 volume using the RWB method on the Intel Paragon (64 nodes) required about 1/3 the time needed on a 200 MHz R44OO SGI workstation. We anticipate that optimizing the code for individual processors will further improve performance. Our initial scalability studies indicate that an 1/0 bottleneck may occur during the simultaneous completion and output of results from the nodes. We plan to investigate this problem further. We also plan to generalize our parallelization strategy by developing parallel codes which run on the CrayT3D and networks of workstations as well as the Intel Paragon. As an important part of our goals, we plan to automate the transfer of data from the microscope, the initiation of task execution on scalable parallel systems, and the return of the computed result over the network. We are currently interfacing the Paragon programs to an asynchronous communication environment (ACE), recently developed as part of an NSF National Challenge Grant project to implement remote control of an HVEM and to transparently distribute computationally intensive processing to high performance computers on the network. We demonstrated a preliminary version of this interface at Supercomputing 95 where we performed a tomographic reconstruction remotely using ACE to initiate processing of data on the Intel Paragon (Young et al., 1996). Young, S. J., Fan, G. Y., Hessler, D., Lamont, S., Elvins, T. T., Hadida, M., Hanyzewski, G., Durkin, J. W., Hubbard,P., Kindalman,D.. Wong, E., Greenberg,D. Karin, S. and Ellisman, M. H. """"""""Implementing a Collaboratory for Digital Microscopy"""""""". Int. J. Supercomputing Applications (in press).
Showing the most recent 10 out of 270 publications
