%0 Journal Article %A Guillem Pratx %A Suleman Surti %A Craig Levin %T GPU-accelerated list-mode reconstruction for 3-D time-of-flight (TOF) PET %D 2009 %J Journal of Nuclear Medicine %P 469-469 %V 50 %N supplement 2 %X 469 Objectives The incorporation of TOF information in the reconstruction of PET data leads to improvements in the image quality, particularly when list-mode fully-3D ordered-subset expectation-maximization (OSEM) is used. However, this algorithm requires large amounts of computation. Hence, large clusters of computers are used for reconstructing TOF datasets. This can be avoided by using an inexpensive graphics processing unit (GPU) to accelerate the projection (back and forward) of lines through the image volume. Methods A new line projection method based on the GPU was adapted to include TOF information. Two Gaussian kernels are used in the projection: one varies with the distance to the LOR axis, the other with the TOF difference. For evaluation, a 35 cm-diameter water-filled cylinder containing six 10 mm-diameter spheres (activity concentration 6:1) was acquired for 5 min (125M counts) on a Gemini TF PET scanner. A 288x288x90 voxels volume was reconstructed on the GPU using list-mode 3D-OSEM (15 iterations, 20 subsets). Attenuation correction (CT-based) and normalization were incorporated in the reconstruction. No other correction was performed. Results GPU-based reconstruction took 7.3 sec per million events processed for TOF and 11.7 sec for non-TOF. The contrast v.s. noise trade-off showed that the contrast recovery coefficient (CRC) at 15 iterations was 10.4% with TOF vs 5.7% without (+83%). Noise was 30% higher with TOF (TOF: 16.5%, non-TOF: 12.7%). Lack of randoms and scatter correction reduced the CRC for both datasets. Conclusions Accelerated 3D list-mode TOF reconstruction can be performed on a GPU. TOF information yielded an increase in SNR. Adding corrections will permit a more quantitative evaluation. Research Support This work was supported by the National Institutes of Health (NIH) under grants R01CA119056, R33EB003283, R01CA120474 and the Stanford Bio-X graduate fellowship program. %U