HOME HELP FEEDBACK SUBSCRIPTIONS ARCHIVE SEARCH TABLE OF CONTENTS		QUICK SEARCH:   [advanced] Author: Keyword(s): Year:  Vol:  Page:

This Article Services Email this article to a friend Similar articles in this journal Alert me to new issues of the journal Download to citation manager Citing Articles Citing Articles via Google Scholar Google Scholar Articles by Pratx, G. Articles by Levin, C. Search for Related Content PubMed Articles by Pratx, G. Articles by Levin, C.

Accelerated list-mode 3D-OSEM reconstruction for PET on a graphics processing unit

¹ Radiology, Stanford University, Stanford, California

526

Objectives: We implemented an accelerated version of list-mode 3D ordered subset expectation maximization (OSEM) on a graphics processing unit (GPU). GPUs, present in every computer, are mainly used to run the computational graphical operations required by video games rendering. GPUs now outperform CPUs in term of performance: the 3GHz Pentium IV raw computation power is 6 billion floating point operations per second (FLOPS), compared to 200 GFLOPS for the Nvidia GeForce 7800 GPU. Because they also offer a better price/performance ratio, GPUs are becoming popular in scientific computing. In the reported work GPUs are used to accelerate high-resolution PET image reconstruction.

Methods: The list-mode OSEM algorithm was implemented on the Nvidia GeForce 6600 GPU, using OpenGL and Cg. We used graphics execution units called shaders to run assembly code directly on the graphics card. Data comprising hot spheres in a warm background were generated using the Monte-Carlo package GATE and reconstructed with one iteration of list-mode 3D-OSEM. Execution time and correctness of the reconstruction were compared to a CPU implementation. No normalization was performed.

Results: Our initial implementation accelerated list-mode 3D-OSEM by a factor of 2 without significant optimization. Images reconstructed with the GPU were less noisy due to slight processing differences in the reconstruction. We expect an acceleration of 10 after optimizations to our implementation are made.

Conclusions: In the past ten years, GPU performance has been increasing at a much faster rate than CPU performance. Our implementation will be extremely useful for clinical and research radionuclide imaging, as it will allow fast and inexpensive reconstruction of both measured and simulated emission tomography data. These results are promising showing that a single, cutting-edge GPU can replace an order-of-magnitude more expensive computer cluster for list mode 3D-OSEM reconstruction.

Research Support (if any): This work was supported in part by grants R21 EB003283 from NIH-NIBIB and R21 CA098691 from NIH-NCI.