MAS 2012 Proceeding

Ultra-fast registration of 2D electron microscopy images

Authors:   Santiago Garcia, Julio Kovacs, Pablo Chacon

Abstract

Registration of noisy images is a crucial problem in multiple fields. In single-particle electron microscopy a huge number of two-dimensional (2D) image projections of biological molecules must be aligned multiple times to obtain a three-dimensional (3D) reconstruction. The 3D reconstruction quality is substantially determined by the accuracy and speed of the alignment method. Here, we implemented and validated a real-space correlation method based on Fourier-Bessel functions to perform the matching with a single 3D Fast Fourier transform. Our results show that we achieved faster and more accurate 2D alignments than state-of-the-art methods available in current image EM processing packages. The parallel-CUDA implementation achieves speedups ranging from 23 to 300 folds, depending on the angular sampling used. The achieved over-performance provides an excellent background for developing advanced single particle analysis algorithms towards higher resolutions. Besides, the method presented here is fully applicable to any 2D image rigid-body registration problem.

I3M  Scientific Sponsors

I3M  Industrial Sponsors

I3M  Media Sponsors