Parallel implementation of 3-D iterative reconstruction with intra-node image update for the jPET-D4

The use of small crystals in new PET scanners results in large number of line-of-response (LOR). Thus, fully 3D image reconstruction becomes time consuming especially for iterative algorithms. One of the methods to speed-up image reconstruction is by parallel computing with computer cluster. However, as number of computing node increases, the parallel efficiency decreases mainly due to network transfer delay. Network delay is significant particularly for iterative reconstruction algorithms with large subset size, such as RAMLA. In this paper, we propose a way to reduce data transfer between nodes by allowing intra-node image update. In details, few intra-node updates for L-1 number of sub-iterations are performed in each slave thread before a global update is done at the master thread. We implemented a normal parallel version and the proposed intra-node update version of RAMLA for the jPET-D4 scanner by using MPICH2 library. We evaluated it with four workstations, each with two dual-core opteron 2.8 GHz processors. There are total of 16 slave threads. Each slave thread calculates different set of LORs which are divided according to ring difference numbers. We accessed image quality of the proposed method with a hotspot phantom generated by GATE simulator. With subset size of 48, log likelihood is compared between normal RAMLA and intra-node update version of L=4, 8, and 16. The difference in log likelihood is negligible even for L=16. Other than simulation, we also reconstruct a Hoffman phantom data from FDG experiment. No image artifact is observed for simulation as well as experiment images. In our implementation, computation time is reduced by 15% with L=16. More significant saving in calculating time is expected with larger subset size by optimizing the relaxation parameter. With intra-node update, we showed the possibility to speed up computation without much deterioration in the image quality.