A challenge in using on-board cone beam computed tomography (CBCT) to image lung tumor motion prior to radiation
therapy treatment is acquiring and reconstructing high quality 4D images in a sufficiently short time for practical use.
For the 1 minute rotation times typical of Linacs, severe view aliasing artifacts, including streaks, are created if a
conventional phase-correlated FDK reconstruction is performed. The McKinnon-Bates (MKB) algorithm provides an
efficient means of reducing streaks from static tissue but can suffer from low SNR and other artifacts due to data
truncation and noise. We have added truncation correction and bilateral nonlinear filtering to the MKB algorithm to
reduce streaking and improve image quality. The modified MKB algorithm was implemented on a graphical processing
unit (GPU) to maximize efficiency. Results show that a nearly 4x improvement in SNR is obtained compared to the
conventional FDK phase-correlated reconstruction and that high quality 4D images with 0.4 second temporal resolution
and 1 mm<sup>3</sup> isotropic spatial resolution can be reconstructed in less than 20 seconds after data acquisition completes.
In modern medical CT, the primary source of data is a set of X-ray projections acquired around the object, which are
then used to reconstruct a discrete regular grid of sample points. Conventional volume rendering methods use this
reconstructed regular grid to estimate unknown off-grid values via interpolation. However, these interpolated values may
not match the values that would have been generated had they been reconstructed directly with CT. The consequence can
be simple blurring, but also the omission of fine object detail which usually contains precious information. To avoid
these problems, in the method we propose, instead of reconstructing a lattice of volume sample points, we derive a highfidelity
object model directly from the reconstruction process, fitting a localized object model to the acquired raw data
within tight tolerances. This model can then be easily evaluated both for slice-based viewing as well as in GPU 3D
volume rendering, offering excellent detail preservation in zooming operations. Furthermore, the model-driven
representation also supports high-precision analytical ray casting.
SC829: MIC-GPU: High-Performance Computing for Medical Imaging on Programmable Graphics Hardware (GPU)
Advanced graphics boards have become a standard ingredient in any mid-range and high-end PC, and aside from enabling stunning interactive graphics effects in computer games, their rich programmability allows speedups (over CPU-based code) of 1-2 orders of magnitude also in general-purpose computations. This course explains, in gentle ways, how to exploit this powerful computing platform to accelerate various popular medical imaging applications, such as CT, MRI, image processing, and data visualization. It begins by introducing the basic GPU architecture and its programming model, which establishes a solid understanding on how general computing tasks must be structured and implemented on the GPU to achieve the desired high speedups. Next, it examines a number of standard 2D and 3D medical imaging operators, such as filtering, sampling, statistical analysis, transforms, projectors, etc, and explains how these can be effectively accelerated on the GPU. Finally, it puts this all together by describing the full GPU-accelerated computing pipeline for a representative set of medical imaging applications, such as analytical and iterative CT, MRI, image enhancement chains, and volume visualization.