Abstract
Iterative reconstruction of high-resolution cone-beam CT data is still a difficult task due to the demand for vast amounts of computer cycles and associated memory. In order to improve the performance of iterative algorithms for cone-beam CT reconstruction, an acceleration approach integrating GPU acceleration, empty space skipping and multi-resolution technique is proposed. The approach divides the reconstructed volume into equally sized blocks, and empty blocks are identified by reconstructing an initial low-resolution volume and segmenting it with threshold method. Then all non-empty blocks are packed into a new volume, which is initialized by interpolating the low resolution volume and reconstructed at full resolution using iterative algorithms. Finally these non-empty blocks are rearranged to get the reconstructed high-resolution volume. The whole process is implemented in parallel based on GPU. Since only the voxels in non-empty blocks are calculated, the number of considered voxels is greatly reduced, which translates directly into substantial computation, memory requirements and data transfer savings. The method is evaluated by reconstructing images from simulated projection data of phantom and CT datasets. The results indicate that our approach significantly improves the performance of iterative reconstruction while maintaining a high image quality, compared to conventional GPU-based approaches.