Implementation and Evaluation of an Exact Cone-Beam Reconstruction Algorithm for Flat Detector CT

Abstract

In this thesis, the Defrise-Clack (DC) algorithm [4], an exact cone-beam reconstruction algorithm, was implemented in CUDA for the most common geometry used in x-ray C-arm CT systems (flat panel detector geometry). The implementation utilizes the forward gridding method with Kaiser-Bessel kernel to calculate the 2D Radon transform efficiently. Furthermore, a complete image quality evaluation (resolution, noise and artifacts) was carried out on reconstructions from 29 different cone-beam data sets and with 196 different sets of reconstruction parameters (Ns,Nμ,W). The results were compared to the predominantly used cone-beam reconstruction algorithm, Feldkamp-Davis-Kress (FDK) [9].To improve reconstruction speed, an alternative implementation of the Defrise-Clack algorithm with inverse gridding method was suggested. The inverse gridding method was originally designed to replace the 2D inverse Radon transform [24], however, we showed that this method, with a small modication and without additional processing, can also replace the 2D backprojection operation. The Defrise-Clack algorithm was then implemented in CUDA with the inverse gridding method and the same image quality evaluations were carried out. The inverse gridding implementation utilizes the same Kaiser-Bessel kernel with the same width parameter W as the forward gridding implementation. Results have shown that this method does not degrade image quality if the size of the Kaiser-Bessel kernel is large enough (W > 2). Results indicate that the Defrise-Clack algorithm provides reconstructions with signicant improvements over the short-scan FDK algorithm with respect to cone-beam artifacts. In addition, it was shown that both algorithms have the sameresolution-noise characteristics.