RESEARCH ARTICLE


Cuda Parallel Implementation of Image Reconstruction Algorithm for Positron Emission Tomography



Belzunce MA1, 2, *, Verrastro CA1, 2, Venialgo E1, 2, Cohen IM1, 3
1 Departamento de Electrónica, Facultad Regional Buenos Aires, Universidad Tecnológica Nacional. Medrano 951, (C1179AAQ) Ciudad Autónoma de Buenos Aires, Argentina
2 Instrumentación y Control, Comisión Nacional de Energía Atómica. Presbítero Luis González y Aragón Nro.15, (B1802AYA) Ezeiza, Buenos Aires, Argentina
3 Facultad Regional Avellaneda, Universidad Tecnológica Nacional. Av Mitre 750, (B1870AAU) Avellaneda, Buenos Aires, Argentina


© 2018 Belzunce et al.

open-access license: This is an open access article distributed under the terms of the Creative Commons Attribution 4.0 International Public License (CC-BY 4.0), a copy of which is available at: https://creativecommons.org/licenses/by/4.0/legalcode. This license permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.

* Address correspondence to this author at the Centro Atómico Ezeiza, Comisión Nacional de Energía Atómica; Tel: (+54) (11) 6779 8277. Fax: (+54) (11) 6779 8433. E-mail: martin.a.belzunce@gmail.com


Abstract

Although the use of iterative algorithms for image reconstruction in 3D Positron Emission Tomography (PET) has shown to produce images with better quality than analytical methods, they are computationally expensive. New Graphic Processor Units (GPUs) provide high performance at low cost and programming tools that make it possible to execute parallel algorithms in scientific applications. In this work, a GPU parallel implementation of the iterative reconstruction algorithm MLEM 3D has been developed using CUDA, a parallel model from NVIDIA. The Siddon algorithm was used as Projector and Backprojector. Acceleration factors up to 85 times were achieved, with respect to a single thread CPU implementation. The performance in GPU with Tesla and Fermi, which are respectively the first and the last generation of CUDA compatible architectures, has been compared. The image quality in each platform has been analyzed, showing a higher level of noise in GPU, due to race condition problems. The new features of Fermi architecture permitted to solve this problem using atomic operations.

Keywords: PET, Iterative image reconstruction, Graphics Processing Units, parallelization, CUDA.