62
reads

Cheng-Ying Chou

2013-03-27
16:35:00 - 17:00:00

101 , Mathematics Research Center Building (ori. New Math. Bldg.)

Image reconstruction methods are essential to applications of many medical imaging modalities. Iterative reconstruction methods can potentially mitigate the effects of data noise or incompleteness, and hence facilitate patient dose reduction, but are not currently suitable for routine clinical use due to their long reconstruction times. The system matrix that relates the sought-after image to the measurement data can be enormous. The commodity graphics hardware possesses a highly parallel computational architecture of the computing units and high bandwidth memory bus, which represents a fast computational resource for medical image reconstruction. To demonstrate the use of a graphics processing unit (GPU) in medical image reconstruction, we take advantage of the GPU features to accelerate the reconstruction algorithms of computed tomography and positron emission tomography. Specifically, we showed how the GPU-based algorithm can be optimized by (i) employing a tiled algorithm with multiple-level parallelization, (ii) optimizing thread block size, (iii) maximizing data reuse on constant memory and shared memory, (iv) exploiting built-in texture memory interpolation capability to increase efficiency, and (v) incorporate the symmetry properties of the imaging system with the GPU architecture, etc. The results showed that the GPU-accelerated scheme can drastically reduce the reconstruction time and thus can largely expand the applicability of these image modalities.