Statistical Reconstruction Methods for 3D Imaging of Biological Samples with Electron Microscopy

Abstract

Electron microscopy has emerged as the leading method for the in vivo study of biological structures such as cells, organelles, protein molecules and virus like particles. By providing 3D images up to near atomic resolution, it plays a significant role in analyzing complex organizations, understanding physiological functions and developing medicines. The 3D images representing the electrostatic potential distribution are reconstructed by utilizing the 2D projection images of the target acquired by electron microscope. There are two main 3D reconstruction techniques in the field of electron microscopy: electron tomography (ET) and single particle reconstruction (SPR). In ET, the projection images are acquired by rotating the specimen for different angles. In SPR, the projection images are obtained by analyzing the images of multiple objects representing the same structure. Then, the tomographic reconstruction methods are applied in both methods to obtain the 3D image through the 2D projections.<br/><br/>Physical and mechanical limitations can prevent to acquire projection images that cover the projection angle space completely and uniformly. Incomplete and non-uniform sampling of the projection angles results in anisotropic resolution in the image plane and generates artifacts. Another problem is that the total applied dose of electrons is limited in order to prevent the radiation damage to the biological target. Therefore, limited number of projection images with low signal to noise ratio can be used in the reconstruction process. This affects the resolution of the reconstructed image significantly. This study presents statistical methods to overcome these major challenges to obtain precise and high resolution images in electron microscopy.<br/><br/>Statistical image reconstruction methods have been successful in recovering a signal from imperfect measurements due to their capability of utilizing a priori information. First, we developed a sequential application of a statistical method for ET. Then we extended the method to support projection angles freely distributed in 3D space and applied the method in SPR. In both applications, we observed the strength of the method in projection gap filling, robustness against noise, and resolving the high resolution details in comparison with the conventional reconstruction methods. Afterwards, we improved the method in terms of computation time by incorporating multiresolution reconstruction. Furthermore, we developed an adaptive regularization method to minimize the parameters required to be set by the user. We also proposed the local adaptive Wiener filter for the class averaging step of SPR to improve the averaging accuracy.<br/><br/>The qualitative and quantitative analysis of the reconstructions with phantom and experimental datasets has demonstrated that the proposed reconstruction methods outperform the conventional reconstruction methods. These statistical approaches provided better image accuracy and higher resolution compared with the conventional algebraic and transfer domain based reconstruction methods. The methods provided in this study contribute to enhance our understanding of cellular and molecular structures by providing 3D images of those with improved accuracy and resolution.