Towards real-time tomography: Fast reconstruction algorithms and GPU implementation

Synchrotron X-ray tomographic microscopy is a powerful technique which allows fast non-destructive, high resolution, quantitative volumetric investigations on diverse samples. Highly brilliant X-rays delivered by third generation synchrotron facilities coupled with modern detector technology permit routinely acquisition of high resolution tomograms in few minutes, making high throughput experiments a reality and bringing real-time tomography closer. Despite the continuous progress in computing technology, fast post-processing of such large amount of data is, however, still difficult, if not impossible, with standard approaches. New solutions are mandatory to fully exploit advantages provided by the high acquisition speed. In this paper, we focus on two different solutions regarding in particular the reconstruction process. On one hand we will discuss a fast reconstruction algorithm, based on the Fourier Transform method. The critical step of such method, the polar-to-Cartesian mapping in the Fourier space, is performed, in the used algorithm, by convolution with the Fourier transform of functions with particular characteristics (1D Prolate Spheroidal Wave Functions). This convolution approach combines speed with accuracy, making real-time data post-processing closer to reality. On the other hand, we will also discuss new acceleration possibilities for standard Filtered Back-Projection algorithms offered by the emerging Graphics Processing Unit (GPU) technology. These solutions reduce the reconstruction time for a 2048x2048 slice (and 1501 projections) down to about 0.5 s on a single CPU (Fourier Transform method) and 2 s on a single GPU (Filtered Back-Projection algorithm).