Model-based 3-D reconstruction of transparent objects from a single image

This dissertation describes an interdisciplinary research project that spans computer-vision, imaging of biological nanoparticles via cryogenic transmission electron microscopy (cryo-tem), and bio-informatics / proteomics. This research introduces a new method for reconstructing individual biological particles imaged via cryo-tem with typical accuracies of 88% to 98%. This individual particle reconstruction (IPR) technique is based on a hidden Markov model derived from the underlying biological structure. The processing presented in this dissertation reconstructs each individual particle from a single image, which is in contrast to current approaches that can require tens of thousands of images to create an ensemble representation of a particle class. The presented approach also has an internal intermediate result structure that provides confidence measures and can lead to updates of the model based upon further statistical analysis. As this interdisciplinary research represents a new area of reconstruction processing, this abstract serves as both an executive summary of the research and as an introduction to concepts that may be unfamiliar to practitioners of one (or more) of the disciplines. The main computer-vision focus of this dissertation presents the IPR algorithm, which is a new model-based technique for reconstructing a transparent object from a single image. Model-based 3-D object reconstruction algorithms operating on 2-D projection images require a model that provides the basis for understanding the representation of the imaged object in the scene. Existing cryo-tem reconstruction approaches rely upon averaging of classified images to reduce noise. Improvements in these iterative approaches occur by omitting those images that do not meet the latest reconstruction classification constraints. Another approach uses statistical methods in order to determine a most probable fit to averaged images of a particle class\u27 members. But, again, this processing reconstructs an ensemble representation. The IPR processing presented here uses a higher-order hidden Markov model schema, with levels of states and durations in the states, to define relationships between the image and the candidate reconstruction object in three dimensions without the need for image averaging. IPR processing also allows for a non-iterative solution that grows only linearly in time with reconstruction resolution. The bio-informatics / proteomics component of this research provides insights into the requirements for models used in the IPR processing. Features of the model arise from the underlying biological structure of the object of interest. These biological structure features include items such as information about the molecular composition, which is obtained by orthogonal and routine analytical techniques. Given their importance to health issues and their basic 3-part composition, low density lipoprotein (LDL) particles were used as the basis family of transparent nanoparticles. The processing approach at the core of this research uses a current biological hypothesis of the details of the LDL macro-assembly structure to derive the higher-order HMM schema. A simplified initial approach to modeling the composition of the LDL particles used the three distinct states {phospholipid head groups, lipids/cholesterol, and protein} as well as a fourth state for the buffer solution. These states were used because it is well known that these components exist in quite different electron density and/or localization regimes. After initial trials, the algorithm was further refined by including the duration that each of the multiple parallel Markov chains stayed in each state as each chain traversed through the particle model. IPR processing of model sizes versus target sizes was performed over the expected radius range for LDL particles. These trials resulted in voxel to voxel matching accuracies greater than 88% from a single 2-D image, even with typical cryo-tem image noise added to each target\u27s simulated projection image. More importantly, these reconstructions matched the actual target geometry parameters within a pixel of the actual values. This accuracy implies that a histogram of these parameters (e.g. radius, height, etc.) could be generated to glean a better understanding of the particle structures. It is expected that the results of this research can be used to help produce statistical data for a population of real images of LDL particles. The ultimate goal for this research is that correlations of these LDL sizes and shapes to related known health markers could lead to an improved diagnostic strategy to predict the risk of coronary heart disease / arteriosclerosis. And, while not directly addressed, the innovative method described here could have other uses for analyzing orthographic projection images of transparent objects. These uses range from general proteomic studies and quality assurance for drug delivery nanoparticles to dental x-rays and other medical and manufacturing inspection imaging opportunities. (Abstract shortened by UMI.