Ultrafast image reconstruction of a dual-head PET system by use of CUDA architecture

Positron emission tomography (PET) is an important imaging modality in both clinical usage and research studies. For small-animal PET imaging, it is of major interest to improve the sensitivity and resolution. We have developed a compact high-sensitivity PET system that consisted of two large-area panel PET detector heads. The highly accurate system response matrix can be computed by use of Monte Carlo simulations, and stored for iterative reconstruction methods. The detector head employs 2.1 x 2.1 x 20 mm 3 LSO/LYSO crystals of pitch size equal to 2.4 mm, and thus will produce more than 224 millions lines of response (LORs). By exploiting the symmetry property in the dual-head system, the computational demands can be dramatically reduced. Nevertheless, the tremendously large system size and repetitive reading of system response matrix from the hard drive will result in extremely long reconstruction times. The implementation of an ordered subset expectation maximization (OSEM) algorithm on a CPU system (four Athlon x64 2.0 GHz PCs) took about 2 days for 1 iteration. Consequently, it is imperative to significantly accelerate the reconstruction process to make it more useful for practical applications. Specifically, the graphic processing unit (GPU), which possesses highly parallel computational architecture of computing units can be exploited to achieve a substantial speedup. In this work, we employed the state-of-art GPU, NVIDIA Tesla C2050 based on the Fermi-generation of the compute united device architecture (CUDA) architecture, to yield a reconstruction process within a few minutes. We demonstrated that reconstruction times can be drastically reduced by using the GPU. The OSEM reconstruction algorithms were implemented employing both GPU-based and CPU-based codes, and their computational performance was quantitatively analyzed and compared.