Fast optimal decomposed modification of FDK with potential decreasing of memory consuming

We present a new algorithm for 3D cone-beam tomography. The algorithm is based on decomposition of the cone-beam backprojection operation and angular decimation. It has computational complexity of $O(N^{3.5})$ and allows considerable reduction of peak memory usage in comparison with conventional algorithms. Tests with real data demonstrate the acceleration, achieved by using our algorithm instead of FDK, with 20-fold speedup for a $2000 \times 2000 \times 720$ image. The algorithm is compared with other fast FDK algorithms.