Fast, Adaptive Expectation-Maximization Alignment for Cryo-EM

Cryo-EM is a method for reconstructing 3D structure of proteins without crystallization. The Expectation-Maximization (EM) algorithm is used in the alignment step of Cryo-EM reconstructions. The EM step is often a serious computational bottleneck for 3D reconstructions. This paper proposes a computationally adaptive version of the EM algorithm that speeds up the algorithm by a factor of 20 - 30. Experiments with noisy real-world data are included to show that the algorithm achieves this speedup without any significant loss of accuracy. Such speed ups are significant, allowing the reconstruction to converge in cpu-days rather than cpu-months.