DM-SIRT: A Distributed Method for Multi-tilt Reconstruction in Electron Tomography

Abstract

The ‘missing wedge’ of single tilt in electron tomography introduces severely artifacts into the reconstructed results. To reduce the ‘missing wedge’ effect, a widely used method is ‘multi-tilt reconstruction’, which collects projections using multiple different axes. However, as the number of tilt series increases, its computing and memory costs also rises. While the demand to speed up its reconstruction procedure grows, the huge memory requirement from the 3D structure and strong data dependencies from projections heavily limit its parallelization. In our work, we present a new fully distributed multi-tilt reconstruction framework named DM-SIRT. To improve the parallelism of the reconstruction process and reduce the memory requirements of each process, we formulate the multi-tilt reconstruction as a consensus optimization problem and design a distributed multi-tilt SIRT algorithm. To improve the reconstruction resolution, we applied a multi-agent consensus equilibrium (MACE) with a new data division strategy. Experiments show that along with the visually and quantitatively improvement in resolution, DM-SIRT can acquire a 5.4x speedup ratio compared to the raw multi-tilt reconstruction version. It also has 87% decrease of memory overhead and 8 times more scalable than the raw reconstruction version.