Based on its region-of-interest (ROI) reconstruction advantage, backprojection-filter algorithm has been used in cone-beam CT recently. However, because of its complexity and computation, there is memory insufficiency in the implementation of GPU acceleration. Hence, CUDA-based parallel implementation for BPF algorithm was proposed. Meanwhile, an accelerated projection scheme and other accelerated techniques were included as well, such as features of BPF for acceleration. Besides, video memory pool was introduced to optimize the implementation. With an efficient structure, the simulation results show that it takes only 8.055 seconds by using the new structure to reconstruct 512×512×512 data and only 4.566 seconds for the ROI reconstruction. The output of first block data takes only 1.523 seconds. With a great decrease of memory occupation from 2.5 GB to less than 100 MB, the new scheme is suitable for big data reconstruction.