Tomographic reconstruction in O(N2 log n) using GPU's

Abstract

This project will develop a tool for image reconstruction from projections, using data obtained from the micro-tomography imaging beamline at the Brazilian Synchrotron Light Source. Such techniques include post-processing of a large dataset, with an average size of 30Gb. The process of a tomographic acquisition should be modular, allowing to detect possible points of improvement or correction in the scientific case addressed. In this project, we will study and implement a part of this flow, presenting and proposing new analytical mathematical methods (or iterative) to reconstruct images from a limited number of angles, and using the Fourier transform as a main source of acceleration, thus reducing the computational effort. Currently, we have a formula with low computational complexity given by O(N2 logN), improving the O(N3) from conventional algorithms. In this work, we will take advantage of using a modern graphics card to process a large volume of data using the new formula, with low cost. Such techniques have great appeal in reducing the x-ray dose in the sample. We will provide implementation of these algorithms for beamline users in programming language Python and CUDA. This work is part of a prototype for future GPU projects which will be integrated at Sirius - the new Brazilian synchrotron light source - where we seek fast reconstruction with high quality. (AU)