A scalable multiprocessor implementation of the reprojection algorithm for volumetric PET imaging

Abstract

Widespread adoption of true volumetric PET imaging requires a fast implementation of a 3D reconstruction algorithm. The clinical environment requires prompt reconstruction of patient images for evaluation. Even in the research setting, where such immediate feedback may not be required, the large number of 3D data frames generated by gated or rapid dynamic protocols mandates a fast reconstructor; one day's scanning needs to be reconstructed in one day's time. We have implemented the Kinahan and Rogers reprojection algorithm on the GE Advance PET scanner with i860 processors in Mercury's RACE multicomputer platform. This implementation supports 64/spl times/64, 128/spl times/128, and 256/spl times/256 matrix sizes; can be stopped and restarted to support algorithm development; is scalable in number of processors; and reconstructs a frame of 128/spl times/128 images in 7:54 with 10 processors, 4:24 with 20 processors, and all corrections applied.