Acceleration of an image reconstruction algorithm for Positron Emission Tomography using a Graphics Processing Unit

por Thomas Felder (IFIC Valencia)

miércoles 25 feb. 2009 12:00 → 13:00 Europe/Madrid

Sala Seminarios IFIC (Edf. Institutos de Investigación)

In this work we applied an iterative image reconstruction algorithm to measured data of a Siemens Biograph16 PET scanner. Furthermore the reconstruction algorithm was also implemented on a graphics processing unit (GPU) to reduce the calculation time. Positron emission tomography (PET) is a medical imaging modality to study metabolic processes in a human or animal body. To reconstruct images from measurement data of a Siemens Biograph16 PET scanner we applied the maximum likelihood - expectation maximization (ML-EM) algorithm. The system matrix, which is needed for the ML-EM, is calculated using the Siddon's ray-tracing algorithm. Basically two approaches were implemented; the online version saves the pre-calculated system matrix to the hard disk and reads the values during computation of the reconstruction process. The online version computes the elements of the system matrix during the reconstruction. The online version was implemented on the graphic card and on the CPU. We evaluated if a medium range graphic card can already accelerate a reconstruction algorithm versus a CPU implementation. The same algorithms were applied to the GPU and the CPU, however, due to the limited and specific hardware the implementation on the graphic card had to be adapted and parallelized. For the GPU applications NVIDIA's programming language compute unified device architecture (CUDA) was used. The reconstructed images showed artifacts that are typical for iterative reconstruction algorithms, as well as background noise. Apart from these effects the images were good in quality. In the GPU application we detected a loss of data due to simultaneous memory accesses in the back-projection process. The error depended, among others, on the resolution of the image. We took various measures that reduced the error, so that we finally achieved images for several resolutions that are comparable in quality to the images calculated by the CPU. We implemented two GPU applications, both run faster than the CPU approaches; the maximum acceleration factor that we achieved is 3.