high performance computing on graphics processing units: hgpu.org

We present an on-line list-mode image reconstruction system using GPUs for a surgical PET imaging probe system. We used the nVidia GeForce 9800GTX+ and CUDA to reconstruct images. The proposed system can generate a three-dimensional image from simulated data in 70 msec. We also compared the processing time with respect to the number of LORs per subset. We are working on optimizing the CUDA code to reduce the processing time. The data for image reconstruction were obtained from Monte Carlo simulations. The PET imaging probe system provides intermediately reconstructed 3-dimensional images online in order to detect small tumors or remnants of surgically removed tumors either during surgery or preoperatively. The on line image reconstruction requires high speed computation. Even multi-processor computers may not satisfy the requirement. Fortunately, currently available GPUs provide high speed computation. We are exploring a variant of one-pass list-mode OSEM in conjunction with GPU-based coding to provide real time image reconstruction. We present intermediately reconstructed images by consecutively applying a non-overlapping data window or an overlapping data window on data. The PET probe system itself consists of a segment of a conventional PET scanner and a small, high-resolution PET detector equipped with a position-tracker. Coincidence lines-of-response (LORs) are collected between the two detectors. The imaging probe’s proximity to the lesions and the small size crystals in the probe contribute to high spatial resolution. The continuously updated 3D image is reprojected in real-time onto a plane whose orientation is driven by the tracking system. Based on the intermediate images and the preoperative PET scans, surgeons can survey suspicious regions in detail. If online reconstruction and display are achievable, the PET imaging probe can be viewed as a handheld, clinician-guided camera capable of seeing the distribution of the radiotracer. Note that the probe suf–fers from an inherent limited angle tomography problem. However, because we are interested in the FOV that is close to the PET imaging probe, the effect of limited angle tomography may not be so severe. The GPU based image reconstruction system will be integrated with a prototype PET imaging probe. While computation speed is not yet sufficient, it can be improved by using high-end GPUs and optimized CUDA code.