ML-EM Implementation on a GPU: Avoiding Simultaneous Read-Modify-Write Processes

Iterative image reconstruction algorithms for positron emission tomography (PET) became more and more common in the last decade. The reconstruction time of these computational intensive approaches can be reduced using graphics rocessing units (GPU). We implemented the ML-EM algorithm to reconstruct measurement data of a Biograph Senstation 16 PET/CT scanner (Siemens). To build the system matrix the Siddon’s algorithm was used. The implementation of the backprojection operation on the GPU showed a possible data loss due to a simultaneous read-modify-write process of parallel threads. In this work we analyze the problem and show that this data loss can lead to worse image quality when the probability of the simultaneous memory access increases. We have developed several strategies on the GPU; a straight forward implementation, one that reduces the probability and one that avoids simultaneous access completely applying atomic functions. Our fastest GPU implementation is 33 faster than the CPU reconstruction.     «

Iterative image reconstruction algorithms for positron emission tomography (PET) became more and more common in the last decade. The reconstruction time of these computational intensive approaches can be reduced using graphics rocessing units (GPU). We implemented the ML-EM algorithm to reconstruct measurement data of a Biograph Senstation 16 PET/CT scanner (Siemens). To build the system matrix the Siddon’s algorithm was used. The implementation of the backprojection operation on the GPU showed...     »