Principles for an Implementation of a Complete CT Reconstruction Tool Chain for Arbitrary Sized Data Sets and Its GPU Optimization.

Jürgen Hofmann,Alexander Flisch,Robert Zboray

doi:10.3390/jimaging8010012

Jürgen Hofmann, Alexander Flisch + Show 1 more

Open Access

https://doi.org/10.3390/jimaging8010012

Copy DOI

Abstract

This article describes the implementation of an efficient and fast in-house computed tomography (CT) reconstruction framework. The implementation principles of this cone-beam CT reconstruction tool chain are described here. The article mainly covers the core part of CT reconstruction, the filtered backprojection and its speed up on GPU hardware. Methods and implementations of tools for artifact reduction such as ring artifacts, beam hardening, algorithms for the center of rotation determination and tilted rotation axis correction are presented. The framework allows the reconstruction of CT images of arbitrary data size. Strategies on data splitting and GPU kernel optimization techniques applied for the backprojection process are illustrated by a few examples.

Highlights

At Empa’s Center for X-ray Analytics, different types of Computed Tomography (CT) scanners are in use
TIGRE, ASTRA and RTK providing iterative solvers, which are capable of handling reconstruction problems that cannot be solved by filtered backprojection
The benchmarks were performed on a stack of images with 2048 pixel width, 2048 pixel height and 1800 projections with the following geometry parameters: SCD of 188.0 mm, SDD of 1017.34 mm and a detector pixel size of 0.2 mm

Summary

Introduction

At Empa’s Center for X-ray Analytics, different types of Computed Tomography (CT) scanners are in use. We implemented the backprojection, the most time-consuming component, on a graphics processing unit (GPU) using NVIDIA’s CUDA toolkit [1]. This allows fast—a few minutes—backprojection for large-volume data sets (≈20483 voxels). TIGRE, ASTRA and RTK providing iterative solvers, which are capable of handling reconstruction problems that cannot be solved by filtered backprojection. These are, e.g., limited angle CT or the integration of physical models in the reconstruction procedure. Our implementation running under Windows is available as open source at: https://github.com/JueHo/CT-Recon, accessed on 15 December 2021

Methods

Results

Conclusion