Parallel software for 3D SPECT imaging based on the 2D+1 approximation of collimator blur

We present the parallel software for the reconstruction of SPECT data based on the '2D+1' model. The '2D+1' model, introduced as approximation of 3D model in order to reduce its recostruction time without degrading the accuracy, has been demostrated to be computationally effective, even for sequential architectures. It produces approximately the same accuracy of the 3D model, with a timing requirement comparable to the one of the 2D approach. The basic idea of the 2D+1 model is decoupling the axial blur from the blurring in planes perpendicular to the z-axis. In this way the axial blur-free data sets can be reconstructed independently, by means a two-dimensional (21)) iterative algorithm. Last, the whole 3D reconstruction is obtained by simply stacking the individual 2D reconstructions. Parallelism can be easily introduced into the '2D+1' model through the solution of the 2D subproblems concurrently.

Performance results, obtained running the parallel code on 32 Pentium II nodes of a Beowulf system connected by a switch with a peak bidirectional bandwith of 100 Mb/sec, confirm the intrinsic parallelism of the '2D+1' approximation.