GPU Acceleration Of Separable Footprint Forward And Back Projection For X-ray CT Image Reconstruction

Iterative image reconstruction methods for X-ray CT can improve image quality and reduce radiation dosages over conventional reconstruction methods,such as filtered back projection(FBP).However,iterative image reconstruction methods require much longer computation time.Most iterative image reconstruction methods require one forward projection and one back projection in each iteration.These operations are the primary computational bottleneck in iterative reconstruction methods.The separable footprint(SF)forward and back projection technique simplifies the calculation of intersecting volumes of image voxels and finite-size beams in a way that is both accurate and efficient for parallel implementation.Wu and Fessler implemented the SF forward and back projector on GPU with NVIDIA's CUDA environment,but didn't achieve satisfying speedup.We proposed a new method to accelerate the SF forward and back projection on GPU with NVIDIA's CUDA environment.The proposed method parallelized over all detector cells,traced image voxels which contribute to each detector cell for the forward projection and parallelized over all 3D image voxels for the back projection.Furthermore,we used shared memory to optimize the memory usage in the forward projection and used multiple GPUs to further accelerate the proposed method.We tested the CUDA implementations on NVIDIA Tesla K80 GPUs.The simulation results have shown that the proposed method is 1.7 times faster than the acceleration method of the SF projectors proposed by Wu and Fessler for both the forward and back projection,and the simulation results of multiple GPUs have shown that the computation time of the proposed method is reduced approximately proportional to the number of GPUs.We also tested the proposed method on two NVIDIA Tesla K80 GPUs and original CPU implementation on two 14-core Intel Xeon E5-2695 v3 CPUs.The simulation results have shown that the proposed method is1.3 times faster than the original CPU implementation for the forward projection and 2.1 times faster for the back projection respectively.