Dynamic Load Balancing on Distributed Listmode Time-of-Flight Image Reconstruction

A major obstacle in performing listmode reconstruction in PET imaging is the increased computation time compared to a conventional frame or histogrammed reconstruction. To overcome this challenge in a clinical setting, it is desirable to distribute the reconstruction task to multiple nodes. A previous work investigated the impact of high performance communication networks and focused mainly on static distribution. In practice, optimal static load balancing is difficult. Therefore we have developed a dynamic load balancing approach, which is flexible and can easily be adapted to a varying number of nodes, and the performance of which is not constrained by variation of the load levels of nodes needed for other tasks or by asymmetric network. In this approach, one of the nodes is designated as the distributor, whose task is to partition the events into small chunks and then distribute those chunks to other nodes for processing. Other nodes, which do the actual data processing, are called workers. Each worker requests a new chunk of data to process upon completion of an old one. In case of the OSEM algorithm, when all chunks have been processed in a subset, the workers are synchronized and the image is updated. This image forms the basis for the next subset. This system has been deployed in the Philips GEMINI-TF PET/CT system. For a whole-body patient scan of 150M events, the event processing time with 8 Xeon 3.6GHz dual-processor computers amounts to approximately 9 minutes for 3 iterations.