@inproceedings{oai:repo.qst.go.jp:00054533,
 author = {Tashima, Hideaki and Yoshida, Eiji and Shinaji, Tetsuya and Futada, Haruhiko and Nagata, Takeshi and Haneishi, Hideaki and Yamaya, Taiga and 田島 英朗 and 吉田 英治 and 品地 哲弥 and 羽石 秀昭 and 山谷 泰賀},
 book = {19th REAL TIME CONFERENCE},
 month = {May},
 note = {We are developing the OpenPET, which can provide an open space observable and accessible to the patient during positron emission tomography (PET) measurement. The most attractive and realistic candidate application is in combination with radiotherapy. OpenPET imaging during particle therapy such as carbon beam treatment has the potential to visualize the irradiation field of a patient because positron emitters are produced via nuclear fragmentation
reactions between the irradiated particle and the atomic nuclei of the irradiated tissue. In addition, as a more challenging application, we are focusing on tumor tracking by means of PET, which is conventionally done by gold marker implantation and X-ray imaging. PET imaging normally takes several minutes for data acquisition, data transferring, image reconstruction, and displaying the image. On the other hand, we are aiming at performing acquisition through displaying in less than a second. We should note that there is a limitation of delay due to the accumulation time of list-mode data sufficient for tumor tracking, because the quality of reconstructed images depends on the amount of list-mode data. Therefore, development of the time-delay correction method by the use of supporting device such as a belt sensor is necessary.
In this study, we proposed a real-time imaging system for the OpenPET and implemented on a small OpenPET prototype. For the proposed system, real-time reconstruction system was implemented on graphical processing unit
(GPU) by the use of compute unified device architecture (CUDA). The one-pass 3D list-mode dynamic row-action maximum likelihood algorithm (LM-DRAMA) was employed for the reconstruction algorithm. In the 3D LMDRAMA, the list-mode data were divided into many subsets, and image was updated for each subset to accelerate convergence. The most time consuming processes in the image reconstruction are forward projection and back projection for each list-mode event. The orientations of the list-mode data in a subset were random. Therefore, they were grouped into two classes according to their predominant direction in order to reduce thread divergence.
Conventionally, calculation for each list-mode event was assigned to separate thread. Furthermore, a list-mode event was divided by image slices and processed in parallel in our implementation. In the experiment, we used GeForce GTX580 GPU, which had 512 processor cores. The number of voxels of the reconstructed images was 767684. The processing time for 1,000,000 events of list-mode data was 1.58 s.
In the real-time imaging system, the data transfer control system limits the event counts to be used in the reconstruction step and the reconstructed images are properly intensified by using the ratio of the used counts to the total counts. The
prototype system showed that the real-time monitoring of a moving radioactive source with a frame rate of 2.0 frames per second and delay of 2.1s.},
 pages = {83--84},
 title = {GPU-Accelerated Real-Time Imaging System for the OpenPET toward Tumor-Tracking Radiotherapy},
 year = {2014}
}