Fast GPU-Based Approach to Branchless Distance-Driven Projection and Back-Projection in Cone Beam CT
Date of Award
Fall 2015
Document Type
Thesis
Degree Name
Master of Science (MS)
Department
Electrical and Computer Engineering
First Advisor
Medeiros, Henry
Second Advisor
Frigo, Frederick
Third Advisor
Gilat-Schmidt, Taly
Abstract
Modern CT image reconstruction algorithms rely on projection and back-projection operations to refine an image estimate in iterative image reconstruction. A widely-used state-of-the-art technique is distance-driven projection and back-projection. While the distance-driven technique yields superior image quality in iterative algorithms, it is a computationally demanding process. This has a detrimental effect on the relevance of the algorithms in clinical settings. A few methods have been proposed for enhancing the distance-driven technique in order to take advantage of modern computer hardware. This study explores a two-dimensional extension of the branchless method, which is a technique that does not compromise image quality. The extension of the branchless method is named “pre-projection integration” because it gets a performance boost by integrating the data before the projection and back-projection operations. It was written with Nvidia’s CUDA framework and carefully designed for massively parallel graphics processing units (GPUs). The performance and the image quality of the pre-projection integration method were analyzed. Both projection and back-projection are significantly faster with pre-projection integration. The image quality was analyzed using cone beam CT image reconstruction algorithms within Jeffrey Fessler’s Image Reconstruction Toolbox. Images produced from regularized, iterative image reconstruction algorithms using the pre-projection integration method show no significant artifacts.