Compressed Sensing Methods for Electron Tomography of Cellular Structure

Abstract

Optimization-based image recovery algorithms exploiting sparse representations, such as compressed sensing (CS), are emerging as powerful tools in image denoising and dosage reduction for biomedical imaging. The efficacy of CS recovery for any imaging application hinges on the structure of the images to be reconstructed, the structure of the measurement process, and the structure of image noise. In particular, the utility of CS-ET for undersampled tomogram reconstruction relies on the image satisfying constraints of an a priori image structural model, a dependency, which hints at the theoretical connections between CS and regularization techniques for image recovery. We examine these issues in the context of electron tomography (ET) for membranous cellular organelles, analyzing the ultrastructural sparsity in the datasets, and addressing the structure of the shot noise. Using numerical simulations, we compare reconstructions from the standard tilt series acquisition technique and a randomized tilt-angle variant, against a benchmark CS measurement scheme. To assess reconstruction quality, CS reconstructions were compared with weighted backprojections of two classes of phantom: (1) derived from experimental data, and (2) derived from a generative model that approximates the structure of membrane-bound organelles. We are investigating the application of the CS approach for decreasing acquisition times and electron dose, and for improving the reconstruction quality of electron tomograms recorded from micrometer-thick specimens of neuronal tissues in the scanning transmission electron microscope.This work was supported in part by the Intramural Research Program of NIBIB, NIH.