Ultrafast EM-OSLE for cluster reconstruction in fully 3D PET

For high resolution PET reconstruction, iterative reconstruction methods, such as OSEM, are widely used. However reconstruction time is very long. Authors developed SSP algorithm that improves reconstruction time with order of magnitude. However, it is impossible to perform real-time interactive reconstruction during scan even with SSP and single computer. Therefore, cluster system is required to perform these real-time reconstructions. We have developed channel bonding technique to remedy communication bottleneck, and new algorithm for simultaneous processing of communication and calculation. Projection process is used for Expectation, and also Backprojection for Maximization. Cluster reconstruction is composed of calculations, which are distributed to multiple nodes, and synchronization of calculated result. New cluster reconstruction algorithm will be presented which is using One-Step-Late Expectation concept (we call it as EM-OSLE). EM-OSLE could perform simultaneous calculation and communication, which solves communication bottleneck. For improving the performance of cluster system, we implemented channel-bonding technique with Four gigabit Ethernet cards. This configuration of cluster system with multi-NIC makes bandwidth almost 400MB/sec with 4 NICs. However, communication speed is still bottleneck of reconstruction due to fast CPU and SSP. EM is comprised of E- and M- steps. To perform E-step in duster reconstruction, projection and backprojection should properly be distributed to each node, and image synchronization part is performed by communication. Each node calculates projection and backprojection, and communicates its results with other nodes after calculation. Then, it repeats calculation and communication for iterations. Calculation uses most of system resources, however, communication occupies less resources. Therefore, we can maximize the usage of system resources by changing projection with One-Step-Late images.