UVM Theses and Dissertations
Format:
Print
Author:
Yu, Lingbo
Dept./Program:
Computer Science
Year:
2013
Degree:
PhD
Abstract:
3D electron microscopy is a powerful tool for studying the structures of isolated macromolecules and cellular components. However, the low contrast and high levels of noise limit its strength in revealing high resolution structures. High resolution reconstructions in 3D electron microscopy can be obtained by averaging multiple volumes with identical signals. Multiple volumes can be acquired from several techniques, for example, sub-volumes of electron tomography and multiple reconstructions from random conical reconstruction. The techniques including tilting are the most suitable for investigating structures of heterogeneous samples. However, most tilting techniques can not collect all the data required for a complete 3D reconstructions because the range of the tilt angle is limited in the electron microscope. The multiple volumes need to be first aligned, analyzed for variations, grouped into clusters of similar structures and then averaged within the classes. The high dimensionality and the low signal-to-noise ratio of 3D volumes create a challenging problem. The missing data further complicate the situation since the artifacts caused by the missing data in different orientations can be easily mistaken as the structural differences. The lack of computational tools that account for the missing data has impeded 3D averaging for high resolution structures for long.
This journal-format dissertation contributes two novel algorithms, which are basics tools for 3D averaging to achieve high resolution in 3D reconstructions. The first algorithm, projection based volume alignment (PBVA), is a fast, accurate and robust algorithm for aligning 3D volumes with arbitrary missing data. The second algorithm, probabilistic principal component analysis using expectation maximization algorithm (PPCA-EM), is a multivariate statistical analysis tool of the 3D data set with missing data. Its strength includes extracting the features, presenting the data in a lowerdimensional. feature space, and estimating the missing data. The application of the algorithms demonstrates their effectiveness and their potentials studying biological assemblies.
This journal-format dissertation contributes two novel algorithms, which are basics tools for 3D averaging to achieve high resolution in 3D reconstructions. The first algorithm, projection based volume alignment (PBVA), is a fast, accurate and robust algorithm for aligning 3D volumes with arbitrary missing data. The second algorithm, probabilistic principal component analysis using expectation maximization algorithm (PPCA-EM), is a multivariate statistical analysis tool of the 3D data set with missing data. Its strength includes extracting the features, presenting the data in a lowerdimensional. feature space, and estimating the missing data. The application of the algorithms demonstrates their effectiveness and their potentials studying biological assemblies.