Simultaneous Algebraic Reconstruction Technique for Electron Tomography using OpenCL




Electron tomography (ET) estimates 3D structure of an object from series of 2D transmission images taken from different perspectives. Various methods can be used to obtain the 3D reconstruction.  Expansion of Graphics Processing Units (GPUs) made it possible to explore use of iterative algebraic techniques in ET. To these methods belong Algebraic Reconstruction Technique (ART), Simultaneous Iterative Reconstruction Technique (SIRT), and Simultaneous Algebraic Technique (SART). While ART and SIRT are employed in various SW packages for ET, SART seems to be neglected in this area.  In this thesis we present GPU-based SART implementation modified for ET, called ET SART. After short introduction to ET, we describe mathematical principles of algebraic techniques and present previous work in this area as well as we introduce the basic concepts behind OpenCL and CUDA. We describe main features of ET SART as well as some practical aspects of its implementation.  Finally, we compare ET SART with other methods and present our results. Part of this work is also implementation of wrapper providing uniļ¬ed API for CUDA and OpenCL.